Deep Feature Deblurring Diffusion for Detecting Out-of-Distribution Objects

Aming Wu, Da Chen, Cheng DengICCV 2023 · OpenReview

Abstract

To promote the safe application of detectors, a task of unsupervised out-of-distribution object detection (OOD-OD) is recently proposed, whose goal is to detect unseen OOD objects without accessing any auxiliary OOD data. For this task, the challenge mainly lies in how to only leverage the known in-distribution (ID) data to detect OOD objects accurately without affecting the detection of ID objects, which can be framed as the diffusion problem for deep feature synthesis. Accordingly, such challenge could be addressed by the forward and reverse processes in the diffusion model. In this paper, we propose a new approach of Deep Feature Deblurring Diffusion (DFDD), consisting of forward blurring and reverse deblurring processes. Specifically, the forward process gradually performs Gaussian Blur on the extracted features, which is instrumental in retaining sufficient input-relevant information. By this way, the forward process could synthesize virtual OOD features that are close to the classification boundary between ID and OOD objects, which improves the performance of detecting OOD objects. During the reverse process, based on the blurred features, a dedicated deblurring model is designed to continually recover the lost details in the forward process. Both the deblurred features and original features are taken as the input for training, strengthening the discrimination ability. In the experiments, our method is evaluated on OOD-OD, open-set object detection, and incremental object detection. The significant performance gains over baselines demonstrate the superiorities of our method.

Introduction

Standard object detection models typically operate under a "closed-set" assumption, meaning they are designed to classify objects into a predefined set of categories encountered during training. However, in real-world deployments—such as autonomous vehicles or robotic systems—models frequently encounter Out-of-Distribution (OOD) objects that were never part of the training set. Detecting these unknown objects is a critical safety requirement to prevent the system from making high-confidence but incorrect predictions.

The specific challenge addressed by this research is unsupervised OOD object detection. The goal is to train a model to identify "unknown" objects without having access to any example OOD data during the training phase. Previous methods often synthesized virtual OOD data by adding random noise to features, but this global operation frequently destroys semantic structure, resulting in poor decision boundaries.

Methodology: Deep Feature Deblurring Diffusion (DFDD)

The authors propose Deep Feature Deblurring Diffusion (DFDD), framing OOD detection as a diffusion problem with a unique twist: replacing standard noise-adding steps with a Gaussian blur process.

Forward Feature Blurring

Instead of adding Gaussian noise, the forward process gradually applies Gaussian blur to the feature maps ($x_0 \to x_1 \to \dots \to x_T$) extracted by the backbone network. Unlike noise, blur is a local operation that removes fine details while retaining the overall semantic structure and input-relevant information. After several steps, the resulting feature $x_T$ becomes a "virtual OOD" feature. These features are strategically positioned near the classification boundary, serving as "hard" examples that challenge the model to distinguish between clear ID objects and blurred, uncertain representations.

Reverse Feature Deblurring

To complement the blurring process, a dedicated deblurring model (such as a U-Net) is trained to perform the reverse operation ($x_T \to \dots \to x_0$). By learning to recover lost details, the model becomes more sensitive to the specific textures and structural details that define in-distribution categories. This dual process strengthens the model's overall discrimination ability.

The DFDD framework: Forward blurring synthesizes virtual OOD features near the decision boundary, while reverse deblurring enhances feature discrimination. The right panel shows successful OOD detection results on real-world testing data.

Joint Training and Uncertainty

The model is trained using a combination of standard detection losses and a specialized uncertainty loss. This loss regularizes the detector to produce low OOD scores for original features and high OOD scores for the synthesized blurred features, effectively tightening the decision boundary around known classes.

Experimental Results

The DFDD method was evaluated across three tasks: Unsupervised OOD-OD, Open-Set Object Detection (OSOD), and Incremental Object Detection (IOD), using datasets such as PASCAL VOC, BDD-100k, and MS-COCO.

  • Significant Error Reduction: On the BDD-100k dataset, DFDD reduced the FPR95 (False Positive Rate at a 95% True Positive Rate) by 13.56% compared to the VOS baseline. This indicates a much lower likelihood of mistaking a mystery object for a known category.
  • Improved Accuracy: Unlike some OOD methods that degrade the performance of the original detector, DFDD maintained or slightly improved the mean Average Precision (mAP) for known classes.
  • Versatility: The method proved effective as a plug-and-play enhancement for other frameworks, showing gains in both incremental learning and open-set detection scenarios.

Conclusion

DFDD introduces a novel approach to outlier synthesis by utilizing feature-level deblurring diffusion. By leveraging the mathematical properties of Gaussian blur to create semantically relevant virtual OOD samples, the framework achieves state-of-the-art performance in identifying unknown objects. While the method introduces some computational overhead due to the reverse deblurring process, its ability to improve safety without auxiliary data makes it a significant step forward for robust object detection.

Summary updated

Abstract

To promote the safe application of detectors, a task of unsupervised out-of-distribution object detection (OOD-OD) is recently proposed, whose goal is to detect unseen OOD objects without accessing any auxiliary OOD data. For this task, the challenge mainly lies in how to only leverage the known in-distribution (ID) data to detect OOD objects accurately without affecting the detection of ID objects, which can be framed as the diffusion problem for deep feature synthesis. Accordingly, such challenge could be addressed by the forward and reverse processes in the diffusion model. In this paper, we propose a new approach of Deep Feature Deblurring Diffusion (DFDD), consisting of forward blurring and reverse deblurring processes. Specifically, the forward process gradually performs Gaussian Blur on the extracted features, which is instrumental in retaining sufficient input-relevant information. By this way, the forward process could synthesize virtual OOD features that are close to the classification boundary between ID and OOD objects, which improves the performance of detecting OOD objects. During the reverse process, based on the blurred features, a dedicated deblurring model is designed to continually recover the lost details in the forward process. Both the deblurred features and original features are taken as the input for training, strengthening the discrimination ability. In the experiments, our method is evaluated on OOD-OD, open-set object detection, and incremental object detection. The significant performance gains over baselines demonstrate the superiorities of our method. The source code will be made available at: https://github.com/AmingWu/DFDD-OOD.
BibTeX
@inproceedings{wu2023deep,
  author = {Aming Wu and Da Chen and Cheng Deng},
  title = {{Deep Feature Deblurring Diffusion for Detecting Out-of-Distribution Objects}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2023},
  url = {https://openreview.net/forum?id=s8ywFGvjaC},
}
Figure