Test-time Correction: An Online 3D Detection System via Visual Prompting
Hanxue Zhang *, Zetong Yang *, Yanan Sun, Li Chen, Fei Xia, Fatma Güney, Hongyang Li · 2024
Read Test-time Correction: An Online 3D Detection System via Visual Prompting by Hanxue Zhang *, Zetong Yang *, Yanan Sun, Li Chen, Fei Xia, Fatma Güney, Hongyang Li (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-07768.txt.
Description
Hanxue Zhang and five co-authors present Test-time Correction, in a 2024 preprint, a system that fixes a self-driving car's 3D detector while it drives. Normal detectors are frozen after training, so their mistakes repeat. This system watches for risky objects and corrects errors on the spot, with no retraining. Help comes from side signals: clashes with 2D detections, road descriptions, or a user's click. An add-on module turns those signals into visual prompts, which are stored in a buffer for later frames.
The prompts steer the detector toward the objects it tends to miss. Across many tests, the system repaired errors far better than the frozen detectors it wrapped. Gains held with few labels, in unseen settings, and in bad weather. The authors hope the work starts a line of post-deployment repair tools. Cars, they argue, should keep learning from their own near misses.
Catalog record
| CREATOR | Hanxue Zhang *, Zetong Yang *, Yanan Sun, Li Chen, Fei Xia, Fatma Güney, Hongyang Li |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 75 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-07768