Simple Supervision Is Hard to Beat: A Bitter Lesson from Sparse Target Labels in Domain-Adaptive Object Detection
Lijun Zhang, Ruinian Xu, Mudit Agrawal · 2026
Read Simple Supervision Is Hard to Beat: A Bitter Lesson from Sparse Target Labels in Domain-Adaptive Object Detection by Lijun Zhang, Ruinian Xu, Mudit Agrawal (2026) 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-2606-30795.txt.
Description
Lijun Zhang, Ruinian Xu and Mudit Agrawal revisit a practical question in object detection: what if a few target images could be labelled after all? Source-free domain adaptation retrains a detector on a new, unlabelled domain using its own pseudo-labels. The team ask how much a small, randomly sampled set of true labels helps, at budgets from one to ten percent.
Their method, Random-Target Supervised Mixing, adds a supervised loss on those labelled images and leaves the unlabelled adaptation branch untouched. Across four adaptation methods, two detectors and many tasks, it improved source-free results consistently, by up to about eighteen points of average precision. The team then tested ten plugins that feed the labels into self-training, through pseudo-label selection, object completion or optimisation control. Those gave only limited, method-dependent extra gains. Their stated lesson: simple supervision is hard to beat.
Catalog record
| CREATOR | Lijun Zhang, Ruinian Xu, Mudit Agrawal |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 74 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.30795