Debiased Novel Category Discovering and Localization

TL;DR

The paper tackles novel category discovery and localization in open-world object detection by addressing detector bias toward seen objects. It introduces Debiased Region Mining (DRM) with a dual RPN that jointly optimizes class-aware localization for knowns and class-agnostic localization for unknowns, coupled with semi-supervised instance-level contrastive finetuning and minibatch K-means clustering to discover and cluster new categories. The method achieves state-of-the-art performance on NCDL benchmarks by improving both localization accuracy and the quality of discovered unknown categories, while maintaining computational efficiency. This approach advances practical open-world detection by enabling simultaneous localization and clustering of novel objects without requiring exhaustive re-labeling.

Abstract

In recent years, object detection in deep learning has experienced rapid development. However, most existing object detection models perform well only on closed-set datasets, ignoring a large number of potential objects whose categories are not defined in the training set. These objects are often identified as background or incorrectly classified as pre-defined categories by the detectors. In this paper, we focus on the challenging problem of Novel Class Discovery and Localization (NCDL), aiming to train detectors that can detect the categories present in the training data, while also actively discover, localize, and cluster new categories. We analyze existing NCDL methods and identify the core issue: object detectors tend to be biased towards seen objects, and this leads to the neglect of unseen targets. To address this issue, we first propose an Debiased Region Mining (DRM) approach that combines class-agnostic Region Proposal Network (RPN) and class-aware RPN in a complementary manner. Additionally, we suggest to improve the representation network through semi-supervised contrastive learning by leveraging unlabeled data. Finally, we adopt a simple and efficient mini-batch K-means clustering method for novel class discovery. We conduct extensive experiments on the NCDL benchmark, and the results demonstrate that the proposed DRM approach significantly outperforms previous methods, establishing a new state-of-the-art.

Figures (5)

• Figure 1: For current mainstream object detection models, they do not detect new objects in unlabeled data. The existing NCDL methods tend to extract biased proposals for unlabeled data. Our proposed model not only localizes unknown objects but also clusters different unknown objects, thereby achieving the discovery of new categories.
• Figure 2: Our Model Architecture. The category discovery and localization pipeline includes three necessary components: 1) learning a feature extractor, 2) discovering boxes from unlabelled images, 3) grouping and categorizing the extracted boxes.
• Figure 3: The effect of three kinds of RPN. The object detector in (a) is used a class-aware RPN. The object detector. In (b) uses an object detector that removes the classification head and learns objectness alone. In (c) is our proposed object detector based on IOU regression.
• Figure 4: We validate our method on the COCO dataset and present the results of its visualization (in the visualization results we remove the boxes with too low confidence). It can be clearly seen that our method can effectively localize more target objects than the standard Faster R-CNN. Compared with OLN-Box, our method is more accurate in locating known classes.
• Figure 5: The impact of the two clustering methods on AuC with varying sample sizes and total number of clusters.