Core Concepts
Incorporating objectness scores enhances the detection of unknown objects in driving scenarios.
Abstract
The article discusses the importance of detecting unknown traffic obstacles for safe autonomous driving. It introduces a method that combines object-detection principles with pixel-wise anomaly detection to achieve stable performance in identifying unknown objects. The proposed approach utilizes a semantic-segmentation network with a sigmoid head to provide objectness scores and unknown scores simultaneously. By integrating these scores, the method defines unknown objectness scores, which outperform state-of-the-art methods in publicly available datasets. The study also explores pixel-wise anomaly detection methods, uncertainty estimation, and the use of OoD data for supervision. Experimental results demonstrate the effectiveness of the proposed method in reducing false-positive predictions, especially in background regions, and improving detection performance.
I. INTRODUCTION
- Detection of unknown traffic obstacles crucial for safe autonomous driving.
- Standard object-detection methods struggle with unknown objects.
- Pixel-wise anomaly detection approach gaining research attention.
- Study aims to achieve stable performance in detecting unknown objects.
II. RELATED WORK
- Two approaches for pixel-wise anomaly detection: autoencoder-based and uncertainty-based.
- Autoencoder-based methods use encoder-decoder modules for reconstruction.
- Uncertainty-based methods rely on uncertainty estimation for anomaly detection.
III. PROPOSED METHOD
- Semantic-segmentation network with a sigmoid head used for objectness and unknown scores.
- Unknown objectness scores defined by combining objectness and unknown scores.
- Loss function introduced to reduce false positives in boundary regions.
IV. EXPERIMENTS
- Evaluation conducted on publicly available datasets with driving scene images.
- Proposed method outperforms state-of-the-art approaches.
- Effectiveness of objectness scores demonstrated in reducing false positives.
V. LIMITATIONS
- Method's performance dependent on objectness scores.
- Future research direction includes strategies for recognizing road obstacles as objects.
VI. CONCLUSIONS
- Novel method presented for identifying unknown objects in driving scenarios.
- Semantic-segmentation network with objectness scores enhances detection performance.
- Potential to improve safety in autonomous driving systems.
Stats
"The proposed method outperforms state-of-the-art methods when applied to the publicly available datasets."
"Quantitative evaluations show that the proposed method outperforms state-of-the-art methods when applied to the publicly available datasets."
Quotes
"The proposed anomaly score—unknown objectness score—reduces the number of false-positive predictions, especially in the background regions."
"The proposed method exhibits better performance than state-of-the-art approaches when applied to the publicly available datasets."