Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Revisiting Out-of-Distribution Detection in Real-time Object Detection: From Benchmark Pitfalls to a New Mitigation Paradigm

Created by
  • Haebom

Author

Changshun Wu, Weicheng He, Chih-Hong Cheng, Xiaowei Huang, Saddek Bensalem

Outline

This paper argues that addressing the overconfidence problem of deep learning-based object detection models on Out-of-Distribution (OoD) inputs requires a comprehensive rethinking of the development lifecycle, beyond algorithmic improvements such as improving the existing scoring function and adjusting test-time thresholds. We highlight the errors (up to 13% contamination) of existing OoD detection evaluation benchmarks and propose a novel training-time mitigation paradigm that fine-tunes the detector using a semantically similar OoD dataset, without relying on external OoD detectors. This approach reduces hallucination errors by 91% in the BDD-100K environment for the YOLO model, and demonstrates generalizability to various detection methods, including YOLO, Faster R-CNN, and RT-DETR, as well as to small-shot adaptation.

Takeaways, Limitations

•
Takeaways:
◦
By revealing serious data contamination issues in existing OoD detection evaluation benchmarks, we raise questions about the reliability of existing research results and emphasize the need for more rigorous evaluation criteria.
◦
We present a novel approach to enhance the OoD resistance of the model itself by leveraging OoD data during training time.
◦
A general methodology is presented that can be applied to various object detection models such as YOLO, Faster R-CNN, and RT-DETR, and can be adapted even with small amounts of data.
◦
We demonstrate substantial performance improvements, reducing the hallucination error of the YOLO model by 91% on the BDD-100K dataset.
•
Limitations:
◦
The effectiveness of the proposed method may be limited to specific datasets and models. Additional experiments on diverse datasets and models are needed.
◦
Further analysis is needed on the computational cost and data preparation complexity of the new training time relaxation paradigm.
◦
The 13% contamination rate is for a specific benchmark, and further verification is needed to determine whether it applies equally to other benchmarks.
👍
Made with Slashpage