Online Open-set Semi-supervised Object Detection with Dual Competing Head

TL;DR

This paper proposes an end-to-end online OSSOD framework that improves performance and efficiency, and proposes a semi-supervised outlier filtering method that more effectively filters the OOD instances using both labeled and unlabeled data.

Abstract

Open-set semi-supervised object detection (OSSOD) task leverages practical open-set unlabeled datasets that comprise both in-distribution (ID) and out-of-distribution (OOD) instances for conducting semi-supervised object detection (SSOD). The main challenge in OSSOD is distinguishing and filtering the OOD instances (i.e., outliers) during pseudo-labeling since OODs will affect the performance. The only OSSOD work employs an additional offline OOD detection network trained solely with labeled data to solve this problem. However, the limited labeled data restricts the potential for improvement. Meanwhile, the offline strategy results in low efficiency. To alleviate these issues, this paper proposes an end-to-end online OSSOD framework that improves performance and efficiency: 1) We propose a semi-supervised outlier filtering method that more effectively filters the OOD instances using both labeled and unlabeled data. 2) We propose a threshold-free Dual Competing OOD head that further improves the performance by suppressing the error accumulation during semi-supervised outlier filtering. 3) Our proposed method is an online end-to-end trainable OSSOD framework. Experimental results show that our method achieves state-of-the-art performance on several OSSOD benchmarks compared to existing methods. Moreover, additional experiments show that our method is more efficient and can be easily applied to different SSOD frameworks to boost their performance.

Figures (5)

• Figure 1: (a) The data setting of the OSSOD task. (b) 1) The previous OSSOD method trained the model with only labeled data. 2) We first improve the performance by our semi-supervised outlier filtering method but face the error accumulation problem: The mispredicted OODs make the decision boundary expand to misclassify more samples. 3) We further propose the Dual Competing OOD head to alleviate the error accumulation and result in better performance.
• Figure 2: The framework of our method. Top: Our DCO head is added to the detector for filtering OODs in the pseudo-labels during training. We propose the semi-supervised outlier filtering strategy to improve the filtering ability. Bottom-left: Training strategy of our DCO head, the pseudo-labeled ID/OODs are used for training the positive head (Note that wrong pseudo-label exists). We label all the unlabeled instances as OOD for training the negative head. Bottom-right: OOD filtering using the DCO head. Two heads compete with each other to decide on ID or OOD. In this case, dog is the ID class, and cat is the OOD class.
• Figure 3: Left: The error accumulation problem with only one OOD detection head. Right: The principle of our DCO head for preventing the problem. In this case, dog is the ID class, and cat is the OOD class. The dashed line represents the decision boundary.
• Figure 4: (a) Performance under different data combinations. (b) The number of ID (left) and OOD (right) pseudo-labeled boxes per image during training for different heads. Pos and neg denote our positive and negative heads respectively.
• Figure 5: Visualization results of pseudo-labels and related scores from the DCO head (pos: the positive head; neg: the negative head). The instances are predicted as ID (blue) or OOD (orange) by comparing the two scores.