Reconstructed Feature Interpolation for Downsampled Video Bounding Boxes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding technologies face challenges in efficiently distributing object recognition tasks across a communication network due to power and computational constraints at acquisition devices, leading to suboptimal video encoding and increased bandwidth and latency.
Innovation Solution
Techniques for interpolating reconstructed feature data, allowing object recognition tasks to be distributed across a network by optimizing video encoding based on inference data, which includes generating bounding boxes for temporally downsampled video portions and interpolating these boxes to facilitate efficient distribution and encoding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If video data is transmitted at full resolution and frame rate, then image quality is maintained, but bandwidth consumption and latency increase
Solution Approach 1:
The video data is segmented into different resolution components: full-resolution frames and lower-resolution intermediate frames. This segmentation allows selective transmission of data at different quality levels, reducing overall bandwidth consumption while maintaining critical image quality information in the full-resolution frames.
Solution Approach 2:
Different regions of the video data are assigned different quality levels. Full-resolution data is transmitted for important regions (such as regions containing objects of interest), while less critical regions use lower resolution. This local quality differentiation reduces bandwidth consumption while preserving essential image quality.
2Productivity
If object recognition tasks are performed at acquisition devices, then processing speed is improved, but power and computational constraints are exceeded
Solution Approach 1:
The object recognition task is extracted from the acquisition device and relocated to a remote server or cloud-based processing system. This extraction allows the acquisition device to operate within its power and computational constraints while still enabling sophisticated object recognition through distributed computing.
Solution Approach 2:
An intermediary system (remote server or cloud platform) is introduced between the acquisition device and the object recognition process. The acquisition device transmits video data to this intermediary, which then performs the computationally intensive object recognition tasks, allowing the acquisition device to remain within its operational constraints.
3Quantity of substance
If temporally downsampled video is transmitted, then bandwidth consumption is reduced, but image quality in temporally downsampled portions deteriorates
Solution Approach 1:
Full-resolution video frames and lower-resolution intermediate frames are merged during playback to reconstruct high-quality video. The lower-resolution frames provide temporal information while the full-resolution frames provide spatial detail, and their combination achieves both bandwidth efficiency and image quality.
Solution Approach 2:
The system dynamically adjusts the transmission of full-resolution versus lower-resolution frames based on scene complexity and motion. During periods of high motion or important events, full-resolution frames are transmitted more frequently, while during stable scenes, lower-resolution frames suffice, optimizing both bandwidth usage and perceived image quality.
Data Source
AI summary
This disclosure discloses a method of interpolating inference data corresponding to reconstructed feature data. The method comprising: receiving reconstructed feature data, wherein the reconstructed feature data corresponds to video data which has been temporally downsampled, generating bounding boxes for the reconstructed feature data, and interpolating bounding boxes for temporally downsampled portions of the video.