Proto-OOD: Enhancing OOD Object Detection with Prototype Feature Similarity

TL;DR

Proto-OOD tackles OOD detection in object detection by using prototype-based similarity scoring between input embeddings $r$ and category prototypes $p$, enhanced with a contrastive loss and a negative embedding generator to simulate OOD features. The framework computes $E = H(r,p) \cdot s$, where $H(r,p)$ is cosine similarity and $s$ is the predicted ID/OOD score, and uses a threshold $\gamma$ to separate ID from OOD during inference. Experiments show improvements in FPR95 and AUROC on MS-COCO and OpenImages when evaluated with ID data from Pascal VOC or BDD100K, and the authors also propose Protocol_B to provide fairer OOD evaluation. The work highlights the effectiveness of prototype representations for OOD object detection and provides a practical pipeline that can be integrated with existing detectors.

Abstract

Neural networks that are trained on limited category samples often mispredict out-of-distribution (OOD) objects. We observe that features of the same category are more tightly clustered in feature space, while those of different categories are more dispersed. Based on this, we propose using prototype similarity for OOD detection. Drawing on widely used prototype features in few-shot learning, we introduce a novel OOD detection network structure (Proto-OOD). Proto-OOD enhances the representativeness of category prototypes using contrastive loss and detects OOD data by evaluating the similarity between input features and category prototypes. During training, Proto-OOD generates OOD samples for training the similarity module with a negative embedding generator. When Pascal VOC are used as the in-distribution dataset and MS-COCO as the OOD dataset, Proto-OOD significantly reduces the FPR (false positive rate). Moreover, considering the limitations of existing evaluation metrics, we propose a more reasonable evaluation protocol. The code will be released.

Figures (5)

• Figure 1: The feature visualization of Pascal VOC testing set. We employ t-SNE for dimensionality reduction of the feature extractor's outputs, and each color represents a category feature.
• Figure 2: The framework of Proto-OOD. During the training stage, Proto-OOD collects category prototypes from embeddings r. The negative embedding generator(NEG) creates negative embeddings $\textbf{r}'$ by r and p. In the testing stage, p and the similarity module's outputs s are used to calculate E to determine if an object is OOD object. H(r,p) outputs the cosine similarity between the embedding r and the prototypes p.
• Figure 3: The details of negative embedding generator.
• Figure 4: Visualization the outputs of two object detection networks, RT-DETRlv2023detrs and DDETRzhu2020deformable, on OOD dataset images, as well as the detection outputs of Proto-OOD on RE-DETRlv2023detrs and SIRENdu2022siren applied to DDETR zhu2020deformable.
• Figure 5: b displays the predictions of the SAFE object detector. All the predictions are used to calculate matrix under $\textbf{protocol}_A$, including the incorrect predictions. c shows the predictions, which are used to calculate matrix under $\textbf{protocol}_B$.