Every Pixel Matters: Center-aware Feature Alignment for Domain Adaptive Object Detector (2008.08574v1)

Abstract: A domain adaptive object detector aims to adapt itself to unseen domains that may contain variations of object appearance, viewpoints or backgrounds. Most existing methods adopt feature alignment either on the image level or instance level. However, image-level alignment on global features may tangle foreground/background pixels at the same time, while instance-level alignment using proposals may suffer from the background noise. Different from existing solutions, we propose a domain adaptation framework that accounts for each pixel via predicting pixel-wise objectness and centerness. Specifically, the proposed method carries out center-aware alignment by paying more attention to foreground pixels, hence achieving better adaptation across domains. We demonstrate our method on numerous adaptation settings with extensive experimental results and show favorable performance against existing state-of-the-art algorithms.

• The paper introduces a novel method that leverages pixel-wise objectness and centerness for effective feature alignment in domain adaptive object detection.
• It combines global adversarial alignment with a center-aware strategy to precisely focus on discriminative object regions, overcoming noisy background interference.
• Experimental results demonstrate significant mAP improvements over state-of-the-art methods across diverse scenarios such as weather changes and synthetic-real transitions.

Center-aware Feature Alignment for Domain Adaptive Object Detection

The paper "Every Pixel Matters: Center-aware Feature Alignment for Domain Adaptive Object Detector" introduces a novel approach to domain adaptive object detection, an area that addresses the challenge of adapting object detectors to unseen domains with different object appearances, viewpoints, or backgrounds. Traditional methods often utilize image-level or instance-level feature alignment, each with inherent limitations. Image-level alignment may conflate foreground and background pixels, leading to noisy alignment, while instance-level alignment using proposals may be negatively impacted by background noise.

This research proposes an innovative domain adaptation framework that focuses on pixel-level objectness and centerness prediction to achieve more effective feature alignment across domains. The core idea is to conduct center-aware alignment by emphasizing foreground pixels, thereby improving adaptability and performance.

Key Methodology

The proposed framework operates in two main stages:

1. Global Feature Alignment: Utilizes a global discriminator to align image-level features to minimize domain gaps. This is achieved through adversarial learning, encouraging the model to generate domain-invariant features.
2. Center-aware Feature Alignment: Introduces a pixel-wise approach where features are aligned based on their objectness and centerness scores. This method enables the model to concentrate on the most relevant parts of the objects, reducing the risk of confusion from cluttered backgrounds. Distinct from prior methods, it leverages center-aware maps to guide feature alignment on the pixel level.

The integration of these two alignment strategies is shown to be complementary. Global alignment addresses broader image-level variations, while center-aware alignment hones in on the key discriminative parts of objects, mitigating the issues faced by aligning inconsistent foreground and background features.

Experimental Results

The efficacy of this framework is demonstrated through extensive evaluations across various domain adaptation scenarios, including weather change scenarios from Cityscapes to Foggy Cityscapes, synthetic-to-real transitions from Sim10k to Cityscapes, and cross-camera adaptations from KITTI to Cityscapes. Across these benchmarks, the proposed method outperforms existing state-of-the-art algorithms, especially when both global and center-aware alignments are utilized in tandem.

Quantitative results illustrate substantial improvements in mean Average Precision (mAP) over baselines and competing approaches. The adoption of multi-scale alignment further enhances robustness to objects of varying sizes.

Implications and Future Directions

The introduction of pixel-wise centerness and objectness estimation to guide domain adaptation represents a significant advancement in adaptive object detection strategies. The proposed center-aware feature alignment can serve as a foundation for further explorations in domain adaptation techniques that require fine-grained feature refinement.

Looking forward, the incorporation of additional contextual knowledge and semantic information in the alignment process may enhance understanding and interpretation of scene dynamics. Furthermore, adapting this approach for other computer vision tasks, such as semantic segmentation or video object detection, could yield broader applicability and benefits. Researchers and practitioners may build upon this work to refine adaptive systems that generalize seamlessly across varied domain conditions, supporting robust real-world applications like autonomous driving and real-time surveillance.