Abstract
This paper describes a new way to make real-time object detection systems, like those used in self-driving cars, more stable and consistent over time. Currently, there's a dilemma: Using all the historical sensor data (like past camera frames) is accurate but slow. Processing only the current frame is fast but can lead to unstable results, such as objects flickering or bounding boxes jumping (jitter). The new technique solves this by using a recursive system with a "lightweight semantic feedback loop". Instead of processing old, massive sensor data, the system feeds a highly compressed summary of its own previous predictions (like past object locations and confidence scores) back into its current processing step. To prevent the model from repeating its own mistakes, a "noise-injected training protocol" is used to make the model more robust. This approach improves how consistently objects are detected, reducing visual errors like flicker and jitter, while remaining fast enough for applications that require very low latency.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kligvasser, Idan; Rivlin, Ehud; and Intrator, Yotam, "Recursive Semantic Feedback With Noise-Injected Training for Stable Perception", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/9609