Depth-Guided Video Inpainting for Autonomous Driving

Abstract

Disclosed are systems and methods for video restoration for autonomous driving. For example, the method stitches multiple depth frames into a 3D map, wherein one or more objects in the depth frames have been previously removed. The method also projects the 3D map onto a first image frame to generate a corresponding depth map, wherein the first image frame includes a target restoration area. For each target pixel within the target restoration area of ​​the first image frame, based on the corresponding depth map, the method also maps the target pixel within the target restoration area of ​​the first image frame to a candidate pixel in the second image frame. The method also determines a candidate color for filling the target pixel. The method also performs Poisson image editing on the first image frame to achieve color consistency at the boundary of the target restoration area of ​​the first image frame and between the inside and the outside. For each pixel in the target restoration area of ​​the first image frame, the method also traces the pixel into adjacent frames and replaces the original color of the pixel with the average of the colors sampled from the adjacent frames.

Description

Technical Field

[0001] Embodiments of the present disclosure generally relate to autonomous driving. More specifically, embodiments of the present disclosure relate to systems and methods for depth-guided video inpainting for autonomous driving. Background Art

[0002] With the advancement of computing power, multimodal sensing has become increasingly popular in recent years. In particular, in the field of autonomous driving (AD), combining multiple sensors to overcome the shortcomings of individual sensors can provide redundancy for safety. Currently, most self-driving vehicles are equipped with light detection and ranging (LIDAR) and cameras for perception and mapping. In addition, simulation systems have become crucial for the development and validation of AD technologies. Instead of using computer graphics to create virtual driving scenarios, augmented real-world images with simulated traffic flow have been proposed to create photorealistic simulated images and renderings. A key component in this photorealistic simulated image and rendering pipeline is removing moving agents on the road to generate clean background street images. This type of data can be generated using deep learning-based augmentation platforms and video inpainting methods.

[0003] In addition to image sensors, mapping services that display street-level panoramic views in their apps also choose to place depth sensors on their capture vehicles. However, due to privacy concerns, these street view images must be post-processed to blur faces and license plates before being released for public access. Thus, for better privacy protection and clearer street imagery, there is a strong desire to completely remove these agents from the road.

[0004] Significant progress has been made in image restoration in recent years. The mainstream approach uses patch-based methods to complete missing regions by sampling and pasting similar patches from known regions or other source images. This approach has naturally extended to video restoration, where not only spatial coherence but also temporal coherence is preserved.

[0005] The basic idea behind video inpainting is that missing regions / pixels within a frame are observed in some other frames of the same video. Driven by this observation, several state-of-the-art methods use optical flow as a guide to fill in the missing pixels either explicitly or implicitly. They have been successfully applied in different scenarios and achieve seamless inpainting results. However, regardless of whether the flow computation is learning-based or not, it is affected by textureless regions. Furthermore, perspective changes in the video can also degrade the quality of the optical flow estimate. These frame-by-frame flow errors are accumulated when filling in the missing pixels from temporally distant frames, leading to distorted inpainting results.

[0006] The advent of deep learning, and in particular generative adversarial networks (GANs), has provided powerful tools for inpainting. For images, some traditional approaches have formulated inpainting as a conditional image generation problem. Although formulated differently, GAN-based inpainting approaches are fundamentally the same as patch-based approaches, as the spirit remains to find similar textures in the training data and fill in the holes. Therefore, while domain adaptation is not an easy task once the input image comes from a different scene, conventional methods must carefully select their training data to match the domain of the input image. Furthermore, GAN-based approaches share the same problem as patch-based methods: they are poor at handling perspective changes in images.

[0007] Regarding inpainting, the principle is to fill target holes by borrowing appearance information from a known source. This source can be a region other than the hole in the same image, an image from the same video, or an image / video of a similar scene. Reducing the search space for correct pixels is crucial. Existing work can be divided into three categories based on different approaches: propagation-based inpainting, patch-based inpainting, and learning-based inpainting.

[0008] Propagation-based methods extrapolate boundary pixels around holes to complete the image. These approaches are successful for areas of uniform color. However, filling large holes with rich texture variations is difficult. Therefore, propagation-based approaches typically inpaint small holes and scratches in images.

[0009] Patch-based methods, on the other hand, not only look at boundary pixels but also search for similar appearances in other regions / images to complete the missing region. This approach has been extended to the temporal domain of video inpainting. For example, optical flow and color are jointly estimated in the missing region to address temporal consistency issues. In general, patch-based methods are better at handling non-stationary visual data. As the name suggests, patch-based methods rely on reliable pixel matching to copy and paste image patches to the missing region. When pixel matching cannot be robustly achieved (for example, in the presence of large perspective or illumination changes), the inpainting results are problematic.

[0010] As for learning-based inpainting, the success of deep learning techniques has recently inspired its application to image restoration. For example, one state-of-the-art approach adds several feature maps to a new Shepard layer to achieve stronger results than deeper network structures. GaNs were first introduced to generate novel photos. This approach can be easily extended to the inpainting domain by formulating it as a conditional image generation problem. Another state-of-the-art approach proposes a context encoder—a convolutional neural network trained to generate the content of any image region conditional on its surroundings. The context encoder is trained to both understand the content of the entire image and generate plausible hypotheses about missing parts. Yet another state-of-the-art approach uses global and local context discriminators to distinguish real images from fake ones. The global discriminator examines the entire image to ensure its overall consistency, while the local discriminator examines only a small region centered around the intact region to ensure the local consistency of the generated patch. Recently, contextual attention mechanisms have emerged in generative inpainting frameworks, further improving inpainting quality. For video inpainting, effective frameworks specifically designed to exploit redundant information across video frames have been developed. They first synthesize spatially and temporally coherent optical flow fields across video frames and then use this synthesized flow field to guide pixel propagation to fill in missing regions in the video. Summary of the Invention

[0011] Embodiments of the present disclosure provide a computer-implemented method, a non-transitory machine-readable medium, and a data processing system for video restoration.

[0012] In a first aspect, some embodiments of the present disclosure provide a computer-implemented method for video restoration, the method comprising: receiving a plurality of sensor data sets comprising depth frames and image frames; for each depth frame, removing one or more objects from the depth frame, thereby generating a plurality of resulting depth frames without the one or more objects; stitching the plurality of resulting depth frames into a three-dimensional (3D) map; optimizing a camera pose of a first image frame having a target restoration area; and projecting the 3D map onto the first image frame to generate a corresponding depth map.

[0013] In a second aspect, some embodiments of the present invention provide a non-transitory machine-readable medium having instructions stored therein, which, when executed by a processor, causes the processor to perform operations comprising: receiving a plurality of sensor data sets comprising depth frames and image frames; for each depth frame, removing one or more objects from the depth frame, thereby generating a plurality of resulting depth frames without the one or more objects; stitching the plurality of resulting depth frames into a three-dimensional (3D) map; optimizing a camera pose of a first image frame having a target repair region; and projecting the 3D map onto the first image frame to generate a corresponding depth map.

[0014] In a third aspect, some embodiments of the present disclosure provide a data processing system comprising: a processor; and a memory connected to the processor to store instructions, which, when executed by the processor, causes the processor to perform operations comprising: receiving multiple sensor data sets comprising depth frames and image frames; for each depth frame, removing one or more objects from the depth frame, thereby generating multiple resulting depth frames without the one or more objects; stitching the multiple resulting depth frames into a three-dimensional (3D) map; optimizing a camera pose of a first image frame having a target repair area; projecting the 3D map onto the first image frame to generate a corresponding depth map; and for each target pixel within the target repair area of ​​the first image frame, mapping the target pixel within the target repair area of ​​the first image frame to a candidate pixel in a second image frame included in the image frame based on the corresponding depth map, and determining a candidate color for filling the target pixel.

[0015] In a fourth aspect, some embodiments of the present disclosure provide a computer program product, comprising a computer program, which, when executed by a processor, implements the following method: receiving multiple sensor data sets comprising depth frames and image frames; for each depth frame, removing one or more objects from the depth frame, thereby generating multiple resulting depth frames without the one or more objects; stitching the multiple resulting depth frames into a three-dimensional map; optimizing the camera pose of a first image frame having a target repair area; and projecting the three-dimensional map onto the first image frame to generate a corresponding depth map. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Embodiments of the present disclosure are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings in which like reference numerals indicate similar elements.

[0017] Figure 1 is a block diagram illustrating a networked system according to one embodiment.

[0018] Figure 2 is a block diagram illustrating an example of an autonomous driving vehicle according to one embodiment.

[0019] Figures 3A-3B is a block diagram illustrating an example of an autonomous driving system for use with an autonomous vehicle according to one embodiment.

[0020] Figure 4A is a block diagram illustrating an example video inpainting system according to one embodiment.

[0021] Figure 4B is a block diagram illustrating example modules included in a depth map construction module according to one embodiment.

[0022] Figure 4Cis a diagram illustrating an example pipeline of a video inpainting system according to one embodiment.

[0023] Figure 4D is a diagram illustrating an example of point cloud projection and image pixels within a target area according to one embodiment.

[0024] Figure 5 is a diagram illustrating an example image and its corresponding dense depth map according to one embodiment.

[0025] Figure 6 is a diagram illustrating example color candidate selection criteria according to one embodiment.

[0026] Figure 7 is a diagram illustrating an example of another color candidate selection criterion according to one embodiment.

[0027] Figure 8 is a diagram illustrating example color coordination results according to one embodiment.

[0028] Figure 9 is a flow chart illustrating an example method of constructing a depth map according to one embodiment.

[0029] Figure 10 is a flowchart illustrating an example of a method of video inpainting according to one embodiment. DETAILED DESCRIPTION

[0030] Various embodiments and aspects of the present invention will be described with reference to the details discussed below, and the accompanying drawings will illustrate various embodiments. The following description and drawings illustrate the present invention but do not limit the present invention. Many specific details are described to provide a thorough understanding of the various embodiments of the present invention. However, in some instances, in order to provide a concise discussion of embodiments of the present invention, well-known or conventional details are not described.

[0031] Reference in the specification to "one embodiment" or "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the present invention. The appearance of the phrase "in one embodiment" in various places in the specification does not necessarily refer to the same embodiment.

[0032] According to some embodiments, in order to obtain clear street views and photo-realistic simulations in autonomous driving, an automatic video restoration algorithm is described herein that is capable of removing traffic agents from a video and synthesizing missing areas using guidance from depth / point clouds. By constructing a dense 3D graph from a stitched point cloud, the frames within the video are geometrically related via this common 3D graph. In order to fill the target restoration area within the frame, pixels from other frames can be converted to current pixels with correct occlusion. In addition, multiple videos can be fused by 3D point cloud registration, so that the target video can be restored using multiple source videos. The motivation of the embodiments of the present disclosure described herein is to solve the problem of long-term occlusions where the occluded area is never visible in the entire video. Therefore, the embodiments of the present disclosure are novel and first enable the fusion of multiple videos for video restoration in order to solve such long-term occlusion problems.

[0033] According to some embodiments, this document describes methods and systems for video inpainting guided by three-dimensional (3D) graphs in AD scenarios. In one embodiment, deep learning-based methods are avoided, allowing the entire pipeline to run solely on a central processing unit (CPU). This makes it easily generalizable to different platforms and use cases, as it does not require a graphics processing unit (GPU) or domain adaptation of the training data. In some embodiments, given that more and more videos are accompanied by depth data, 3D graph-guided inpainting is a new direction being explored by the inpainting community.

[0034] According to some embodiments, methods for depth-guided restoration of street view videos are described herein, as image + depth sensors become standard for AD vehicles. For example, depending on the task, target objects are manually marked or automatically detected in a color image, and then removed from their depth counterparts. A 3D map can be constructed by stitching all point clouds together and projecting the 3D map back onto a single frame. Most frame pixels are assigned depth values ​​via 3D projection, and those remaining pixels obtain their depth by interpolation. Using a dense depth map and known external camera parameters, colors can be sampled from other frames to fill in holes in the current frame. These colors act as initial guesses for those missing pixels, followed by regularization to enforce spatial and photometric smoothness. Color coordination can then be applied to generate smooth and seamless blending boundaries. Finally, a moving average is applied along the optical flow to make the final restored video smooth in time.

[0035] Unlike learning-based methods, the above approach cannot repair occluded areas if the occluded area has never been visible in the video. To address this issue, a fusion repair that uses multiple video clips to repair the target area can be used. Compared to the repair approaches of the prior art, the embodiments of the present disclosure described herein are able to maintain better details in the missing area with correct perspective distortion. Since the 3D map is consistent in all frames, temporal consistency is implicitly enforced. In addition, multiple video clips captured at different times can be repaired by aligning all frames into a common 3D point map. In one embodiment, a sensor dataset captured or collected from an AD vehicle may serve as input to the embodiments of the present disclosure described herein, although other suitable datasets (such as datasets involving indoor and outdoor scenes) may be utilized, as long as synchronized image + depth data is utilized.

[0036] In one aspect, a computer-implemented method for video inpainting is described. The method may receive multiple sensor datasets comprising depth frames and image frames. For each depth frame, the method may further remove one or more objects from the depth frame, thereby generating multiple resulting depth frames without the one or more objects. The method may further stitch the multiple resulting depth frames into a 3D map. The method may further optimize the pose of a camera for a first image frame having a target inpainting region. The method may further project the 3D map onto the first image frame to generate a corresponding depth map.

[0037] In one embodiment, for each target pixel within the target repair area of ​​the first image frame, based on the corresponding depth map, the method may also map the target pixel within the target repair area of ​​the first image frame to a candidate pixel in the second image frame included in the image frame. The method may also determine a candidate color for filling the target pixel. The method may also perform Poisson image editing on the first image frame to achieve color consistency between the inside and outside of the target repair area of ​​the first image frame. The method may also use video fusion repair to repair occluded areas within the target repair area. For each pixel in the target repair area of ​​the first image frame, the method may track the pixel to adjacent frames and replace the original color of the pixel with the average of the colors sampled from the adjacent frames.

[0038] Figure 1 is a block diagram showing an autonomous driving network configuration according to one embodiment of the present disclosure. Figure 1, a network configuration 100 includes: an autonomous driving vehicle (ADV) 101, which can be communicatively coupled to one or more servers 103-104 via a network 102. Although one ADV is shown, multiple ADVs can be coupled to each other and / or to the servers 103-104 via the network 102. The network 102 can be any type of wired or wireless network, such as a local area network (LAN), a wide area network (WAN) (such as the Internet, a cellular network, a satellite network), or a combination thereof. The servers 103-104 can be any type of server or server cluster, such as a web or cloud server, an application server, a back-end server, or a combination thereof. The servers 103-104 can be data analysis servers, content servers, traffic information servers, map and point of interest (MPOI) servers, or location servers, etc.

[0039] An ADV refers to a vehicle that can be configured in an automated mode, in which the vehicle navigates an environment with little or no input from the driver. Such an ADV may include a sensor system having one or more sensors configured to detect information about the environment in which the vehicle operates. The vehicle and its associated controller use the detected information to navigate the environment. The ADV 101 can operate in a manual mode, a fully automated mode, or a partially automated mode.

[0040] In one embodiment, ADV 101 includes, but is not limited to, an automated driving system (ADS) 110, a vehicle control system 111, a wireless communication system 112, a user interface system 113, and a sensor system 115. ADV 101 may also include certain common components included in a conventional vehicle, such as an engine, wheels, a steering wheel, a transmission, etc., which may be controlled by vehicle control system 111 and / or ADS 110 using various communication signals and / or commands, such as acceleration signals or commands, deceleration signals or commands, steering signals or commands, braking signals or commands, etc.

[0041] Components 110 - 115 may be communicatively coupled to one another via an interconnect, a bus, a network, or a combination thereof. For example, components 110 - 115 may be communicatively coupled to one another via a controller area network (CAN) bus. CAN bus is a vehicle bus standard designed to allow microcontrollers and devices to communicate with one another in applications without a host computer. It is a message-based protocol originally designed for multiplexed electrical wiring within automobiles, but is also used in many other environments.

[0042] Now refer to Figure 2In one embodiment, the sensor system 115 includes, but is not limited to, one or more cameras 211, a global positioning system (GPS) unit 212, an inertial measurement unit (IMU) 213, a radar unit 214, and a light detection and ranging (LIDAR) unit 215. The GPS system 212 may include a transceiver operable to provide information about the ADV's location. The IMU unit 213 may sense changes in the ADV's position and orientation based on inertial acceleration. The radar unit 214 may represent a system that utilizes radio signals to sense objects within the ADV's local environment. In some embodiments, in addition to sensing objects, the radar unit 214 may also sense the object's speed and / or heading. The LIDAR unit 215 may utilize lasers to sense objects in the ADV's environment. Among other system components, the LIDAR unit 215 may include one or more laser sources, a laser scanner, and one or more detectors. The camera 211 may include one or more devices that capture images of the ADV's surroundings. The camera 211 may be a still camera and / or a video camera. For example, the camera may be mechanically movable by mounting the camera on a rotating and / or tilting platform.

[0043] The sensor system 115 may also include other sensors, such as a sonar sensor, an infrared sensor, a steering sensor, a throttle sensor, a brake sensor, and an audio sensor (e.g., a microphone). The audio sensor may be configured to capture sound from the environment surrounding the ADV. The steering sensor may be configured to sense the steering angle of the steering wheel, the vehicle's wheels, or a combination thereof. The throttle sensor and the brake sensor may respectively sense the vehicle's throttle position and brake position. In some cases, the throttle sensor and the brake sensor may be integrated into an integrated throttle / brake sensor.

[0044] In one embodiment, the vehicle control system 111 includes, but is not limited to, a steering unit 201, a throttle unit 202 (also referred to as an acceleration unit), and a brake unit 203. The steering unit 201 is used to adjust the direction or forward direction of the vehicle. The throttle unit 202 is used to control the speed of the motor or engine, which in turn controls the speed and acceleration of the vehicle. The brake unit 203 decelerates the vehicle by providing friction to slow down the wheels or tires of the vehicle. Note that, as Figure 2 The components shown may be implemented in hardware, software, or a combination thereof.

[0045] Return to reference Figure 1The wireless communication system 112 is used to allow communication between the ADV 101 and external systems, such as devices, sensors, other vehicles, etc. For example, the wireless communication system 112 can wirelessly communicate with one or more devices directly or via a communication network, for example, communicating with servers 103 to 104 through the network 102. The wireless communication system 112 can use any cellular communication network or wireless local area network (WLAN), for example, using WiFi to communicate with another component or system. The wireless communication system 112 can directly communicate with devices (e.g., a passenger's mobile device, a display device, a speaker in the vehicle 101) using, for example, an infrared link, Bluetooth, etc. The user interface system 113 can be part of a peripheral device implemented in the vehicle 101, including, for example, a keyboard, a touch screen display device, a microphone, and a speaker.

[0046] Some or all functions of ADV 101 may be controlled or managed by ADS 110, particularly when operating in automated driving mode. ADS 110 includes the necessary hardware (e.g., processor, memory, storage) and software (e.g., operating system, planning and routing programs) to receive information from sensor system 115, control system 111, wireless communication system 112, and / or user interface system 113, process the received information, plan a route or path from a starting point to a destination, and then drive vehicle 101 based on the planned and controlled information. Alternatively, ADS 110 may be integrated with vehicle control system 111.

[0047] For example, a user acting as a passenger may specify a starting location and destination for a trip, for example, via a user interface. ADS 110 obtains data related to the trip. For example, ADS 110 may obtain location and route data from an MPOI server, which may be part of servers 103-104. The location server provides location services, while the MPOI server provides map services and location-specific points of interest (POIs). Alternatively, this location and MPOI information may be cached locally in persistent storage within ADS 110.

[0048] As ADV 101 moves along a route, ADS 110 can also obtain real-time traffic information from a traffic information system or server (TIS). Note that servers 103-104 may be operated by a third-party entity. Alternatively, the functionality of servers 103-104 may be integrated with ADS 110. Based on real-time traffic information, MPOI information, and location information detected or sensed by sensor system 115, as well as real-time local environmental data (e.g., obstacles, objects, nearby vehicles), ADS 110 can plan an optimal route and, for example, drive vehicle 101 according to the planned route via control system 111 to safely and efficiently reach a specific destination.

[0049] Figure 3A and 3B is a block diagram illustrating an example of an autonomous driving system for use with an ADV according to one embodiment. The system 300 may be implemented as Figure 1 The ADV 101 includes, but is not limited to, an ADS 110, a control system 111, and a sensor system 115. Figures 3A-3B The ADS 110 includes but is not limited to a positioning module 301 , a perception module 302 , a prediction module 303 , a decision module 304 , a planning module 305 , a control module 306 , a route arrangement module 307 , and a sensor data acquisition module 308 .

[0050] Some or all of the modules 301-308 may be implemented in software, hardware, or a combination thereof. For example, these modules may be installed in permanent storage 352, loaded into memory 351, and executed by one or more processors (not shown). Note that some or all of these modules may be communicatively coupled to Figure 2 Some or all modules of the vehicle control system 111 or Figure 2 Some or all modules of the vehicle control system 111 are integrated. Some of the modules 301-308 can be integrated together as integrated modules.

[0051] The positioning module 301 determines the current location of the ADV 300 (e.g., using the GPS unit 212) and manages any data related to the user's trip or route. The positioning module 301 (also known as the map and route module) manages any data related to the user's travel or route. A user can log in, for example, via a user interface and specify a starting location and destination for a trip. The positioning module 301 communicates with other components of the ADV 300, such as the map and route data 311, to obtain trip-related data. For example, the positioning module 301 can obtain location and route data from a location server and a map and point of interest (MPOI) server. The location server provides location services, while the MPOI server provides map services and location-specific point of interest (POIs), which can be cached as part of the map and route data 311. As the ADV 300 moves along its route, the positioning module 301 can also obtain real-time traffic information from a traffic information system or server.

[0052] Based on the sensor data provided by the sensor system 115 and the positioning information obtained by the positioning module 301, the perception module 302 determines the perception of the surrounding environment. The perception information may represent what a typical driver would perceive around the vehicle being driven. This perception may include, for example, lane configuration in the form of objects, traffic light signals, the relative position of other vehicles, pedestrians, buildings, sidewalks, or other traffic-related signs (e.g., stop signs, yield signs). Lane configuration includes information describing one or more lanes, such as the shape of the lanes (e.g., straight or curved), the width of the lanes, the number of lanes in the road, whether they are one-way or two-way, merging or diverging lanes, exit lanes, etc.

[0053] Perception module 302 may include a computer vision system or functionality of a computer vision system to process and analyze images captured by one or more cameras to identify objects and / or features in the ADV environment. Objects may include, for example, traffic signs, road boundaries, other vehicles, pedestrians, and / or obstacles. The computer vision system may utilize object recognition algorithms, video tracking, and other computer vision techniques. In some embodiments, the computer vision system may map the environment, track objects, and estimate their speed. Perception module 302 may also detect objects based on other sensor data provided by other sensors, such as radar and / or LIDAR.

[0054] For each object, prediction module 303 predicts how the object will behave in the environment. Prediction is performed based on sensory data regarding the driving environment at that point in time, given a set of map / route information 311 and traffic regulations 312. For example, if the object is a vehicle traveling in opposite directions and the current driving environment includes an intersection, prediction module 303 will predict whether the vehicle is likely to proceed straight ahead or turn. If the sensory data indicates that the intersection lacks a traffic light, prediction module 303 may predict that the vehicle may have to come to a complete stop before entering the intersection. If the sensory data indicates that the vehicle is currently in a left-turn-only lane or a right-turn-only lane, prediction module 303 may predict that the vehicle is more likely to make a left or right turn, respectively.

[0055] For each object, decision module 304 makes a determination about how to handle the object. For example, given a particular object (e.g., another vehicle in an intersection) and its metadata describing the object (e.g., speed, direction, steering angle), decision module 304 determines how to encounter the object (e.g., overtake, yield, stop, pass). Decision module 304 may make this determination based on a set of rules (such as traffic rules or driving rules 312), which may be stored in persistent storage 352.

[0056] The routing module 307 is configured to provide one or more routes or paths from a starting point to a destination. For a given itinerary from a starting location to a destination (e.g., a given itinerary received from a user), the routing module 307 obtains route and map information 311 and determines all possible routes or paths from the starting location to the destination. The routing module 307 may generate a reference line in the form of a topographic map for each route it determines from the starting location to the destination. A reference line is an ideal route or path that is free of interference from other factors, such as vehicles, obstacles, or traffic conditions. In other words, if there are no other vehicles, pedestrians, or obstacles on the road, the ADV should follow the reference line exactly or closely. The topographic map is then provided to the decision module 304 and / or planning module 305. The decision module 304 and / or planning module 305 examines all possible routes to select and modify the optimal route based on other data provided by other modules (such as traffic conditions from the positioning module 301, the driving environment perceived by the perception module 302, and traffic conditions predicted by the prediction module 303). The actual path or route used to control the ADV may be close to or different from the reference line provided by the routing module 307 , depending on the particular driving circumstances at that point in time.

[0057] Based on the decision for each perceived object, the planning module 305 plans a path or route for the ADV using the reference lines provided by the routing module 307 as a basis, as well as driving parameters (e.g., distance, speed, and / or steering angle). That is, for a given object, the decision module 304 decides what to do with the object, while the planning module 305 determines how to do it. For example, for a given object, the decision module 304 may decide to overtake the object, while the planning module 305 may determine whether to overtake to the left or right of the object. Planning and control data is generated by the planning module 305, which includes information describing how the vehicle 300 will move in the next movement cycle (e.g., the next route / path segment). For example, the planning and control data may instruct the vehicle 300 to move at 30 miles per hour (mph) for 10 meters and then change to the right lane at 25 mph.

[0058] Based on the planning and control data, the control module 306 controls and drives the ADV according to the route or path defined by the planning and control data by sending appropriate commands or signals to the vehicle control system 111. The planning and control data includes sufficient information to drive the vehicle from a first point on the route or path to a second point using appropriate vehicle settings or driving parameters (e.g., throttle, brake, steering commands) at different points in time along the route or path.

[0059] In one embodiment, the planning phase is performed over multiple planning cycles (also called drive cycles), for example, in 100-millisecond (ms) time intervals. For each planning cycle or drive cycle, one or more control commands are issued based on the planning and control data. That is, every 100 ms, the planning module 305 plans the next route or path segment, for example, including a target location and the time required for the ADV to reach the target location. Alternatively, the planning module 305 may specify a specific speed, direction, and / or steering angle. In one embodiment, the planning module 305 plans the route or path segment for the next predetermined time period (e.g., 5 seconds). For each planning cycle, the planning module 305 plans the target location for the current cycle (e.g., the next 5 seconds) based on the target location planned in the previous cycle. The control module 306 then generates one or more control commands (e.g., throttle, brake, and steering control commands) based on the planning and control data for the current cycle.

[0060] Note that decision module 304 and planning module 305 can be integrated into an integrated module. Decision module 304 / planning module 305 can include a navigation system or functionality of a navigation system to determine a driving path for the ADV. For example, the navigation system can determine a series of speeds and headings to influence the movement of the ADV along a path that substantially avoids perceived obstacles while generally navigating the road-based path to the final destination. The destination can be set based on user input via user interface system 113. The navigation system can dynamically update the driving path while the ADV is in operation. The navigation system can combine data from a GPS system and one or more maps to determine the driving path for the ADV.

[0061] In one embodiment, as the ADV moves along the route, the sensor data acquisition module 308 may capture or collect sensor data generated from the sensor system 115 (e.g., camera 211, radar unit 214, LIDAR unit 215, etc.), such as camera data, LIDAR data or frames, radar data, etc., and store the sensor data as part of a sensor dataset 313. The sensor dataset 313 may be stored in the persistent storage device 352 or, alternatively, on a server (e.g., servers 103-104) for subsequent processing. In one embodiment, the sensor dataset 313 may comprise a large-scale dataset of video recorded over a period of time. For example, the video may include synchronized images, point clouds of depth images, etc. In some embodiments, the dataset 313 may include challenging scenes, such as background occlusion by large vehicles (such as cars, buses, or trucks) in intersections, or a forward vehicle that consistently blocks the forward view. For these long-term occlusions, the background may be missing throughout the entire video sequence. Therefore, these challenging scenes may be captured more than once to provide data for video fusion inpainting, as described in more detail below.

[0062] Figure 4A is a block diagram illustrating an example video restoration system according to one embodiment. Figure 4A (as well as Figures 4B-4C ), video restoration system 400 includes a depth map construction module 401, a candidate color sampling module 402, a regularization module 403, a color coordination module 404, and a temporal smoothing module 405. In one embodiment, depth map construction module 401 receives (or retrieves) sensor dataset 313, for example, from persistent storage 352 or a server (e.g., servers 103-104). Sensor dataset 313 may include depth frames 510 (e.g., LIDAR frames), where each depth image includes a frame-by-frame point cloud representing one or more objects and a background of the scene. Sensor dataset 313 may also include image frames 560, for example, captured from camera 211 of system 115.

[0063] like Figure 4BAs shown, depth map construction module 401 may include a dynamic object removal module 421, a map stitching module 422, and a pose optimization module 423. For each depth frame 510, dynamic object removal module 421 removes moving objects from the point cloud, retaining only background points in the final 3D map. This operation can be performed relatively easily once calibration between the depth sensor (e.g., LIDAR unit 215) and the image sensor (e.g., camera 211) has been performed. For example, to remove each moving object, all points projected within one or more bounding boxes (or target regions) in the image surrounding its respective object (e.g., bounding boxes 531-532 in image 530) can be removed. Bounding boxes can be automatically detected or manually labeled. Alternatively, machine learning (e.g., PointNet++) can be utilized on the point cloud to detect and remove those moving objects directly from the point cloud.

[0064] Using the resulting point clouds from depth frames 510, map stitching module 422 can stitch those resulting point clouds into a 3D map 520, removing moving objects. For example, in one embodiment, module 422 can invoke odometry and mapping methods or tools (such as the LIDAR Odometry and Mapping (LOAM) method) in real time to fuse the resulting depth frames to construct 3D map 520. In another embodiment, a real-time 3D reconstruction and interactive method can be used to reconstruct 3D map 520 based on the resulting depth frames. This method can also downsample to generate a final point cloud with a reasonable resolution.

[0065] In one embodiment, the pose optimization module 423 is configured to optimize the camera pose for each image (e.g., image 530). For example, the relative pose between the depth sensor and the image sensor may be pre-calibrated. However, there may still be some misalignment between the point cloud and the image pixels. Vibration, inaccurate synchronization, and / or accumulated errors from point cloud stitching may cause pose offsets between the image sensor and the depth sensor. As an example, refer to Figure 4D , the point cloud is projected into a colored target region 610. The image on the left shows the calibration result projection. Clearly, misalignment is visible at the boundary. The image on the right shows the optimized rotation R projection, where the points more effectively match the surrounding pixels. The area between target region 620 and target region 630 is where the colors of the projected 3D points (e.g., the point cloud in 3D image 520) are compared with image pixels (e.g., pixels in image 530) to optimize the camera rotation matrix R.

[0066] In order to produce seamless restoration results, even if this offset is small, it should be compensated. From the initial extrinsic calibration between the image sensor and the depth sensor, their respective rotation R and translation T are optimized by minimizing the photometric projection error. The error is defined as:

[0067] 2，

[0068] where p is the pixel projection from the 3D map, Φ is the region surrounding the target inpainted area, which is Figure 4D q is shown as the region between target region 620 and target region 630, and q is the original pixel in the image covered by p. Function c returns the value of the pixel.

[0069] Note that the color and position of a pixel are discrete values, making the error function E discontinuous over R and T. The above equation cannot be solved directly using standard solvers such as the Levenberg-Marquardt algorithm or the Gauss-Newton algorithm. Instead, the discrete space of R and T is searched to minimize E. However, R and T may have a total of six degrees of freedom (DoF), making the search space very large. Therefore, T can be fixed and only R optimized, since R dominates in determining the projection position when most of the 3D image is distant points. Furthermore, in some embodiments, only the projected pixels need to be moved slightly in the vertical and horizontal directions in the image space, which is determined by the camera (e.g., Figure 2 Finally, the search space can be reduced to 2DoF, which significantly speeds up the optimization process.

[0070] Return to reference Figures 4B-4C Once the camera pose is optimized by module 423, the 3D map 520 is projected onto each image frame (e.g., image 530) to generate a corresponding depth map 540. Note that some point clouds are captured far from the current image and may be occluded and unoccluded during the projection process. Therefore, a z-buffer may be used to obtain the nearest depth.

[0071] To obtain a fully dense depth map, a depth estimation method can be used to learn from the sparse depth map and generate a dense depth map, although linear interpolation may be sufficient to generate a dense 3D map. In some embodiments, a median filter can be applied to remove some individual noise points, and Figure 5 The final example dense depth map is shown in Figure 5 An image and its corresponding dense depth map are shown in [1]. Note that depth is rendered only for background points and all moving objects have been removed.

[0072] Continue to refer Figure 4AIn one embodiment, the candidate color sampling module 402 is configured to map pixels from one image to pixels from other images. When each pixel is assigned a depth value, it is possible to map pixels from one image to other images. There are multiple color choices for the pixels that fill the target repair area, and criteria should be followed to find the best candidate color. For example, there are two principles to select the best candidate color: 1) select from frames that are temporally closer to the current frame, and 2) select from frames where the 3D background is closer to the camera. The first requirement ensures that the video repair approach described in this article is less subject to perspective distortion and occlusion. The second requirement is that because the image records more texture details when it is closer to the object, more details can be retained in the repair area.

[0073] Under this criterion and the fact that the sensor moves only forward during capture, module 402 is configured to first search forward in time to the end of the video and then search backward to the beginning. The first valid pixel is selected as a candidate pixel, and a valid pixel means that its position does not fall into the target inpainting area. Now referring to Figure 6 , Figure 6 An example of candidate selection criteria according to one embodiment is shown in the top row. Pixel 812 of frame 810 can find its candidate color in two subsequent frames, 820 and 830, where road texture clearly appears. In this case, the frame temporally close to the current frame (e.g., frame 810) is selected to minimize the effects of perspective changes and potential occlusion or disocclusion. In the bottom row, pixel 852 of frame 850 can find its candidate color in both the previous frame 840 and the subsequent frame 860. In this case, since road texture is missing in frame 840, subsequent frame 860 is selected over previous frame 840.

[0074] At this point, each pixel is individually assigned a color value. If the camera pose and depth values ​​are 100% correct, a perfect restoration result can be generated with smooth boundaries and neighborhoods. However, this is not the case in the real world, especially since depth maps often carry errors. Therefore, the regularization module 403 is called to enforce the smoothness constraint. In one embodiment, color selection can be formulated as a discrete global optimization problem and solved using, for example, belief propagation (BP). However, before explaining this formulation, the color space and neighborhood of the target pixel are first defined. Now referring to Figure 7 As shown in the left image pair (image 910-image 920), a target pixel 912 in frame 910 finds its candidate pixel 922 from the source image 920. Due to depth inaccuracies, the true color may not lie entirely on the candidate pixel, but rather on a small surrounding window 921. All pixel colors from the n × n window 921 are then collected to form the color space of the target pixel 912, where n is a positive integer.

[0075] Still refer to Figure 7 , the right image pair (image 930-image 940) shows how to find the expected color of the neighborhood. Due to perspective change, the four neighbors 933 of the target pixel 932 are not necessarily the neighbors in the source image 940. Therefore, the neighboring pixels 943 can be warped to their depth values ​​in the source image 940 to sample the expected color.

[0076] In more detail, in one embodiment, let P be the target repair area (e.g., Figure 7 is a set of pixels in the region 911 / 931 of , and L is a set of labels. The labels correspond to the indices of the potential colors in the color space. The labeling function l is used for each pixel p P distribute lp L . It is assumed that the label should vary smoothly almost everywhere, but may change sharply in some places such as object boundaries. The quality of the labeling is given by the energy function:

[0077]

[0078] where N represents the number of edges in the four-connected image grid. V(lp, lq) is the cost of assigning labels lp and lq to two adjacent pixels and can be referred to as the discontinuity cost. DP(lp) is the cost of assigning label lp to pixel p and can be referred to as the data cost. Determining the label with the minimum energy corresponds to the maximum a posteriori (MAP) estimation problem.

[0079] Therefore, the boundary smoothness constraint can be incorporated into the data cost as follows:

[0080] dp (lp) =

[0081] where Cp1, CpR, Cpt, and Cpb return the expected colors of the left, right, top, and bottom neighbors of pixel p, respectively. q is a neighbor pixel of p that is outside the inpainted region in the target image, so it has a known color, which is returned by the function I(q). In one embodiment, the difference between the true and expected neighborhood colors is used as a measure of labeling quality. For pixels that are not on the inpainting boundary, all labels are given an equal chance by assigning a constant value α. The discontinuity cost is defined as:

[0082] v (lp, lq) =

[0083] Here, Cp(·) and Cq(·) capture the colors for p and q at labels lp and lq. L, R, T, and B are the abbreviations for q's position on the left, right, top, and bottom, respectively. For a pair of two adjacent pixels p and q, the difference between p's color and q's expected color is calculated, and vice versa.

[0084] In one embodiment, the color coordination module 404 is configured to use Poisson image editing to generate smooth and seamless blending boundaries. As an example, due to changes in camera exposure time and white balance, pixels from different frames may have different colors, thereby causing color discontinuities (such as in Figure 8 These problems can be addressed using Poisson image editing. Poisson image editing was originally proposed to clone image patches from a source image into a destination image with seamless boundaries and the original texture. It does this by solving the following minimization problem:

[0085] .

[0086] Ω is a bounded Repair area of ​​Ω. f is the color function of the destination image, and f is the color function of the target inpainting area within the destination image. . = [ . / x, . / y] is the gradient operator. v is the desired color gradient defined on Ω.

[0087] Here, v is computed using the output from the belief propagation step, with one exception: if two adjacent pixels within Ω are from different frames, their gradients can be set to 0. This allows for color consistency within the inpainted region. Figure 8 The effectiveness of this solution is demonstrated. Figure 8 In the figure, left image 1010 is the input image. Middle image 1020 is the inpainted result. Note the color discontinuity in region 1022 and the blank pixels in region 1021. Note that blank pixel region 1021 is also filled. Since blank pixels have a gradient value of zero, solving the Poisson equation for this region is equivalent to smooth color interpolation. Finally, right image 1030 is the result after color coordination.

[0088] Here, it is assumed that the repaired area is visible in some other frames. Otherwise, some pixels will remain blank, as shown in Figure 8 Visible in region 1021. Learning-based methods can inpaint colors based on their training data. In contrast, if the occluded region is never visible in the video, the embodiments of the present disclosure described herein cannot inpaint the occluded region, leaving blank pixels.

[0089] For small areas of blank pixels, smooth interpolation is sufficient to fill these holes. However, in some embodiments, for example, a vehicle ahead can block a large field of view for the entire duration of the video, leaving large blank holes. Simple interpolation cannot handle this problem. This problem can be solved by capturing another video of the same scene, in which the occluded portion becomes visible. Fortunately, the LOAM method can be used to register the newly captured frame to the existing 3D map. Once the new frame is registered and merged into the existing 3D map, inpainting is performed in exactly the same manner.

[0090] Continue to refer Figure 4A , the temporal smoothing module 405 can calculate forward and backward optical flows for all result frames. That is, for each pixel in the target inpainting area, the module 405 can use optical flow to track the pixel to adjacent frames and replace the original color of the pixel with the average of the colors sampled from adjacent frames.

[0091] Figure 9 is a flow chart illustrating an example method for constructing a depth map according to one embodiment. The method 900 may be performed by hardware, software, or a combination of both. For example, the method 900 may be performed by Figure 4B The depth map construction module 401 is executed.

[0092] Reference Figure 9 At block 901, the method receives a multi-sensor dataset comprising depth frames and image frames. At block 902, for each depth frame, the method removes one or more objects from the depth frame, thereby generating a plurality of resulting depth frames without the one or more objects. At block 903, the method stitches the resulting depth frames into a 3D map. At block 904, the method optimizes the camera pose of a first image frame having a target inpainting region. At block 905, the method projects the 3D map onto the first image frame to generate a corresponding depth map.

[0093] Figure 10 is a flow chart illustrating an example of a video restoration method according to one embodiment. The method 1000 may be performed by hardware, software, or a combination of both. For example, the method 1000 may be performed by Figure 4A Executed by system 400.

[0094] Reference Figure 10At box 1001, the method stitches multiple depth frames into a 3D map, where one or more objects in the depth frames have been previously removed. At box 1002, the method projects the 3D map onto a first image frame to generate a corresponding depth map, where the first image frame includes a target inpainting region. At box 1003, for each target pixel within the target inpainting region of the first image frame, based on the corresponding depth map, the method maps the target pixel within the target inpainting region of the first image frame to a candidate pixel in the second image frame. At box 1004, the method determines a candidate color for filling the target pixel. At box 1005, the method performs Poisson image editing on the first image frame to achieve color consistency between the inside and outside of the target inpainting region of the first image frame. At box 1006, for each pixel in the target inpainting region of the first image frame, the method traces the pixel to adjacent frames and replaces the original color of the pixel with the average of the colors sampled from the adjacent frames.

[0095] It should be noted that some or all of the components shown and described above may be implemented in software, hardware, or a combination thereof. For example, such components may be implemented as software installed and stored in a permanent storage device, which may be loaded into a memory by a processor (not shown) and executed in the memory to implement the processes or operations throughout the present application. Alternatively, such components may be implemented as executable code programmed or embedded into dedicated hardware (such as an integrated circuit (e.g., an application-specific integrated circuit or ASIC), a digital signal processor (DSP), or a field-programmable gate array (FPGA)), which may be accessed via corresponding drivers and / or an operating system from an application. Furthermore, such components may be implemented as specific hardware logic in a processor or processor core as part of an instruction set that may be accessed by a software component through one or more specific instructions.

[0096] Some portions of the foregoing detailed description have been presented in terms of algorithms and symbolic representations of operations on data bits within a computer memory. These algorithmic descriptions and representations are the means used by those skilled in the data processing arts to most effectively convey the substance of their work to others skilled in the art. An algorithm, as used herein, is generally conceived to be a self-consistent sequence of operations leading to a desired result. These operations are those requiring physical manipulation of physical quantities.

[0097] It should be borne in mind, however, that all of these and similar terms are intended to be associated with the appropriate physical quantities and are merely convenient labels applied to these quantities. Unless expressly indicated otherwise in the foregoing discussion, it should be understood that throughout this specification, discussions utilizing terms such as those set forth in the appended claims refer to the actions and processes of a computer system or similar electronic computing device that manipulates data represented as physical (electronic) quantities within the computer system's registers and memories and transforms that data into other data similarly represented as physical quantities within the computer system's memories or registers or other such information storage, transmission, or display devices.

[0098] Embodiments of the present disclosure also relate to apparatus for performing the operations described herein. Such a computer program is stored on a non-transitory computer-readable medium. A machine-readable medium includes any mechanism for storing information in a form readable by a machine (e.g., a computer). For example, a machine-readable (e.g., computer-readable) medium includes a machine-readable (e.g., computer-readable) storage medium (e.g., read-only memory ("ROM"), random access memory ("RAM"), magnetic disk storage media, optical storage media, and flash memory devices).

[0099] The processes or methods depicted in the aforementioned figures may be performed by processing logic comprising hardware (e.g., circuitry, dedicated logic, etc.), software (e.g., embodied on a non-transitory computer-readable medium), or a combination of both. Although the processes or methods are described above in terms of sequential operations, it should be understood that some of the operations may be performed in a different order. Furthermore, some operations may be performed in parallel rather than sequentially.

[0100] The embodiments of the present disclosure are not described with reference to any particular programming language. It will be appreciated that a variety of programming languages ​​can be used to implement the teachings of the embodiments of the present disclosure as described herein.

[0101] In the foregoing description, embodiments of the present disclosure have been described with reference to specific exemplary embodiments thereof. It will be apparent that various modifications may be made to the present disclosure without departing from the broader spirit and scope of the present disclosure as set forth in the appended claims. Accordingly, the present description and drawings should be read in an illustrative rather than a restrictive sense.

Claims

1. A computer-implemented method for video restoration, comprising: receiving a plurality of sensor data sets including depth frames and image frames; for each of the depth frames, removing one or more objects from the depth frame, thereby generating a plurality of resulting depth frames without the one or more objects; stitching the plurality of resulting depth frames into a three-dimensional image; Optimizing the camera pose of the first image frame having the target repaired region; as well as projecting the three-dimensional image onto the first image frame to generate a corresponding depth map; For each target pixel in the target restoration area of ​​the first image frame, mapping the target pixels within the target restoration area of ​​the first image frame to candidate pixels in a second image frame included in the image frame based on the corresponding depth map, and Warping neighboring pixels around the candidate pixel to the depth values ​​of the neighboring pixels in the second image frame to sample the desired color to fill the target pixel; wherein warping the neighboring pixels around the candidate pixel into the second image frame includes: calculating an energy function based on a set of pixels in the target repair area and a set of labels corresponding to the indices of the candidate colors in the color space, and incorporating a boundary smoothness constraint into a data cost based on the expected colors of each of the neighboring pixels, wherein the data cost refers to the cost of assigning labels to pixels in the energy function.

2. The method according to claim 1, further comprising: Poisson image editing is performed on the first image frame to achieve color consistency between the interior and exterior of the target restoration area of ​​the first image frame.

3. The method according to claim 1, wherein Each of the depth frames includes a point cloud representing one or more objects and a scene background.

4. The method according to claim 2, further comprising: For each pixel in the target repair area of ​​the first image frame, the pixel is traced to adjacent frames, and the original color of the pixel is replaced by an average value of colors sampled from the adjacent frames.

5. The method according to claim 1, wherein The second image frame is temporally close to the first image frame.

6. The method according to claim 1, wherein The second image frame is a previous frame from the first image frame or a subsequent frame from the first image frame.

7. The method according to claim 2, wherein: Performing Poisson image editing on the first image frame includes computing the following minimization function: ， Where Ω is a bounded The target repair area of ​​Ω, f is the color function of the first image frame, f is the color function of the target repair area in the first image frame, . = [ . / x, . / y] is the gradient operator, and v is the desired color gradient defined on Ω.

8. A non-transitory machine-readable medium having stored therein instructions that, when executed by a processor, cause the processor to perform operations comprising: receiving a plurality of sensor data sets including depth frames and image frames; for each of the depth frames, removing one or more objects from the depth frame, thereby generating a plurality of resulting depth frames without the one or more objects; stitching the plurality of resulting depth frames into a three-dimensional image; Optimizing the camera pose of the first image frame having the target repaired region; as well as Projecting the three-dimensional image onto the first image frame to generate a corresponding depth map; For each target pixel in the target restoration area of ​​the first image frame, mapping the target pixels within the target restoration area of ​​the first image frame to candidate pixels in a second image frame included in the image frame based on the corresponding depth map, and Warping neighboring pixels around the candidate pixel to the depth values ​​of the neighboring pixels in the second image frame to sample the desired color to fill the target pixel; wherein warping the neighboring pixels around the candidate pixel into the second image frame includes: calculating an energy function based on a set of pixels in the target repair area and a set of labels corresponding to the indices of the candidate colors in the color space, and incorporating a boundary smoothness constraint into a data cost based on the expected colors of each of the neighboring pixels, wherein the data cost refers to the cost of assigning labels to pixels in the energy function.

9. The non-transitory machine-readable medium of claim 8, wherein: The operation further includes performing Poisson image editing on the first image frame to achieve color consistency between the interior and exterior of the target restoration area of ​​the first image frame.

10. The non-transitory machine-readable medium of claim 8, wherein: Each of the depth frames includes a point cloud representing one or more objects and a scene background.

11. The non-transitory machine-readable medium of claim 9, wherein: The operation further includes: for each pixel in the target repair area of ​​the first image frame, tracing the pixel to adjacent frames, and replacing the original color of the pixel with an average value of colors sampled from the adjacent frames.

12. The non-transitory machine-readable medium of claim 8, wherein: The second image frame is temporally close to the first image frame.

13. The non-transitory machine-readable medium of claim 8, wherein: The second image frame is a previous frame from the first image frame or a subsequent frame from the first image frame.

14. The non-transitory machine-readable medium of claim 9, wherein: Performing Poisson image editing on the first image frame includes computing the following minimization function: ， Where Ω is a bounded The target repair area of ​​Ω, f is the color function of the first image frame, f is the color function of the target repair area in the first image frame, . = [ . / x, . / y] is the gradient operator, and v is the desired color gradient defined on Ω.

15. Data processing system, including: processor; as well as a memory coupled to the processor to store instructions that, when executed by the processor, cause the processor to perform operations comprising: receiving a plurality of sensor data sets including depth frames and image frames; for each of the depth frames, removing one or more objects from the depth frame, thereby generating a plurality of resulting depth frames without the one or more objects; stitching the plurality of resulting depth frames into a three-dimensional image; Optimizing the camera pose of the first image frame having the target repaired region; Projecting the three-dimensional image onto the first image frame to generate a corresponding depth map; and For each target pixel in the target restoration area of ​​the first image frame, mapping the target pixel in the target restoration area of ​​the first image frame to a candidate pixel in a second image frame included in the image frame based on the corresponding depth map, and Warping neighboring pixels around the candidate pixel to the depth values ​​of the neighboring pixels in the second image frame to sample the desired color to fill the target pixel; wherein warping the neighboring pixels around the candidate pixel into the second image frame includes: calculating an energy function based on a set of pixels in the target repair area and a set of labels corresponding to the indices of the candidate colors in the color space, and incorporating a boundary smoothness constraint into a data cost based on the expected colors of each of the neighboring pixels, wherein the data cost refers to the cost of assigning labels to pixels in the energy function.

16. Computer program product comprising a computer program which, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images