Image Depth Model Training With Dynamic-Object Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training deep networks for vehicle image recognition is hindered by poor accuracy due to dynamic object movements in training images, leading to inaccurate depth information recognition.
Innovation Solution
An image depth recognition model is trained using instance segmentation to distinguish static and dynamic objects, selecting target dynamic objects based on pixel points and positions, and generating a target image to filter out moving objects, thereby improving training accuracy by minimizing the impact of dynamic objects on loss values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If training images include dynamic objects, then the training data is more representative of real-world scenarios, but the training accuracy of the deep network deteriorates due to movements of dynamic objects
Solution Approach 1:
The patent applies instance segmentation to divide the training image into multiple instance masks, each corresponding to a specific object. By segmenting dynamic objects from static background, the system can selectively process different regions with appropriate strategies, thereby maintaining representativeness while improving accuracy.
Solution Approach 2:
The patent extracts and removes dynamic objects from the training images by identifying them through instance segmentation and comparing across multiple images. By taking out these moving objects, the training data becomes more suitable for depth recognition while still retaining static scene information.
2Measurement precision
If dynamic objects are removed from training images, then the training accuracy improves, but the training data becomes less representative of real-world scenarios
Solution Approach 1:
The patent implements a dynamic processing approach where the system adaptively identifies and handles dynamic objects based on their detection in the image sequence. Rather than statically removing all dynamic content, the system dynamically adjusts the training process by selectively masking or weighting regions containing moving objects, thus preserving representativeness while improving accuracy.
3Measurement precision
If instance segmentation is performed on all dynamic objects, then the accuracy of depth recognition improves, but the computational complexity and processing time increase
Solution Approach 1:
The patent applies instance segmentation selectively to regions containing dynamic objects rather than processing the entire image uniformly. By focusing computational resources on local regions with moving objects, the system achieves improved depth recognition accuracy while reducing overall computational complexity compared to full-image processing.
Solution Approach 2:
The patent performs instance segmentation on a subset of the image (only regions with dynamic objects) rather than the complete image. This partial action approach achieves the necessary accuracy improvement while avoiding the excessive computational burden of processing every pixel in the entire training image.
Data Source
AI summary
The present application relates to image processing and provides a method for training an image depth recognition model, a method for recognizing image depth, and an electronic device. The method obtains static objects, dynamic objects, a dynamic position by performing an instance segmentation on the first image and the second image. A target dynamic object and a feature dynamic object are selected from the dynamic objects and the dynamic objects. A target image and a target projection image are generated according to the target dynamic object and the feature dynamic object. A depth recognition model is trained based on the target image, and the target projection image. The to-be-recognized image is recognized by the depth recognition model.