Simple supervision beats complex methods in sparse-label object detection
Labeling just 1-10% of images outperforms sophisticated self-training tricks
Domain-adaptive object detection (SFDA-OD) typically uses teacher-student self-training with pseudo-labels to adapt a source-trained detector to an unlabeled target domain. In a new paper, researchers revisit this setting with a small, uniformly sampled set of labeled target images. They introduce RTSM (Random-Target Supervised Mixing), which simply incorporates these few annotations via a supervised detection loss while leaving the original unlabeled adaptation branch unchanged. Evaluations span four SFDA-OD methods, two object detectors (e.g., Faster R-CNN), multiple domain shifts, and label budgets from 1% to 10% — RTSM consistently improves pure SFDA by 1.7 to 18.3 AP50.
The team then tested whether the same annotations could drive further gains by steering the unlabeled self-training process. They evaluated ten sparse-label feedback plugins covering pseudo-label selection, object completion, and optimization control. The results were limited and method-dependent, with no plugin consistently outperforming simple RTSM. The paper calls this a 'bitter lesson': when you have a small number of target labels, complex self-training mechanisms add little value over straightforward supervised learning. RTSM provides a simple yet effective baseline for future work in sparse-label SFDA-OD.
- RTSM boosts AP50 by 1.7 to 18.3 points across four SFDA-OD methods with only 1–10% labeled target images.
- Evaluated on two detectors (Faster R-CNN variants) and multiple domain adaptation tasks, including synthetic-to-real.
- Ten advanced sparse-label plugins failed to consistently beat the simple supervised loss, showing complex approaches are unnecessary.
Why It Matters
Shows that in domain-adaptive object detection, a tiny amount of labeled data with simple supervision outperforms complex self-training methods.