Ambiguity-Resistant Semi-supervised Learning (ARSL) for Single-stage Object Detection

The paper proposes an Ambiguity-Resistant Semi-supervised Learning (ARSL) algorithm for single‑stage semi‑supervised object detection, introducing two generic modules: Joint‑Confidence Estimation (JCE) and Task‑Separation Assignment (TSA). JCE jointly evaluates the quality of pseudo‑labels by combining classification confidence and localization quality, while TSA partitions samples into positive, negative, and ambiguous candidates based on the teacher model’s joint confidence, further selecting potential positives for classification and localization tasks.

Background and Motivation : Deep‑learning based object detectors usually require large annotated datasets. To reduce labeling cost, semi‑supervised object detection (SSOD) leverages a small set of labeled data and abundant unlabeled data, typically using a Mean‑Teacher framework and pseudo‑labeling. However, under this pipeline, single‑stage detectors (e.g., FCOS) gain far less improvement than two‑stage detectors (e.g., Faster R‑CNN).

What Limits Single‑stage Detectors in SSOD? Quantitative analysis reveals two major ambiguities in pseudo‑labels: Selection Ambiguity (mismatch between classification confidence and localization quality) and Assignment Ambiguity (incorrect sample‑to‑box assignments). These ambiguities are more severe in single‑stage detectors because they rely heavily on dense predictions.

Algorithm Overview : ARSL addresses the above issues with:

Joint‑Confidence Estimation (JCE) : A dual‑branch network predicts classification scores and localization quality; their product serves as a joint confidence score. For labeled data, IoU‑based soft labels are used for joint training; for unlabeled data, the teacher’s maximum joint confidence is directly employed.

Task‑Separation Assignment (TSA) : Instead of box‑based assignment, TSA uses the teacher’s joint confidence to split samples into negative, positive, and ambiguous groups via a double‑threshold strategy. Ambiguous samples are further filtered for each task: all are used for classification consistency learning, while only those with high similarity to true positives (in class, geometry, and location) are kept for localization training.

Experimental Results : On COCO‑Standard splits (1%, 2%, 5%, 10% labeled data) ARSL consistently outperforms state‑of‑the‑art SSOD methods, with larger gains when large‑scale jittering is added. On COCO‑Full, ARSL achieves a more significant improvement within a shorter training schedule.

Ablation Studies : The ablations show that the baseline FCOS gains only 4.7% AP under the basic semi‑supervised framework, while ARSL raises it to 36.9% AP (+6.2% AP). JCE contributes ~4.0% AP and TSA ~2.2% AP. Further analysis confirms that JCE filters higher‑quality pseudo‑labels and TSA dramatically increases true positives while reducing false positives.

Links to the paper (https://arxiv.org/abs/2303.14960) and code repository (https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/semi_det) are provided for reproducibility.