SCL: Towards Accurate Domain Adaptive Object Detection via Gradient Detach Based Stacked Complementary Losses

Abstract: Unsupervised domain adaptive object detection aims to learn a robust detector in the domain shift circumstance, where the training (source) domain is label-rich with bounding box annotations, while the testing (target) domain is label-agnostic and the feature distributions between training and testing domains are dissimilar or even totally different. In this paper, we propose a gradient detach based stacked complementary losses (SCL) method that uses detection losses as the primary objective, and cuts in several auxiliary losses in different network stages accompanying with gradient detach training to learn more discriminative representations. We argue that the prior methods mainly leverage more loss functions for training but ignore the interaction of different losses and also the compatible training strategy (gradient detach updating in our work). Thus, our proposed method is a more syncretic adaptation learning process. We conduct comprehensive experiments on seven datasets, the results demonstrate that our method performs favorably better than the state-of-the-art methods by a significant margin. For instance, from Cityscapes to FoggyCityscapes, we achieve 37.9% mAP, outperforming the previous art Strong-Weak by 3.6%.

• The paper introduces a novel stacked complementary loss framework combined with gradient detach that improves unsupervised domain adaptive object detection.
• It enhances feature learning by adding deep supervision from multiple layers with losses tailored for different semantic levels.
• Experiments on seven datasets, including Cityscapes to FoggyCityscapes, showcase a 3.6% mAP improvement over previous methods.

Domain Adaptive Object Detection through Stacked Complementary Losses

The paper "SCL: Towards Accurate Domain Adaptive Object Detection via Gradient Detach Based Stacked Complementary Losses" introduces a novel approach for unsupervised domain adaptive object detection. This task involves training an object detector using a source domain that is rich in labels and annotations, while the target domain, where the detector is applied, lacks such labels. The primary challenge in this scenario arises due to the domain shift—the disparities between feature distributions of source and target domains.

Technical Contributions

The paper proposes a method leveraging "Stacked Complementary Losses" (SCL) combined with a gradient detach training strategy. The inclusion of multiple losses at different stages of the network, along with a primary detection loss, enhances the network's ability to learn discriminative features across domains. The proposed approach addresses key challenges that existing methods may overlook, particularly the interaction and compatibility of different loss functions when optimizing the model.

The method's effectiveness is demonstrated through experiments conducted on seven datasets, achieving superior performance compared to state-of-the-art techniques. For instance, the approach yields a mean Average Precision (mAP) of 37.9% when adapting from Cityscapes to FoggyCityscapes, surpassing the previous record by 3.6%.

Key Ideas

The paper outlines two pivotal concepts: deep supervision through diverse complementary losses and gradient detach training. Deep supervision allows the model to receive feedback from multiple intermediate layers, fostering better feature learning across both domains. The complementary losses are carefully chosen—cross-entropy, least-squares, and focal losses—to match the semantic levels of feature processing. Preceding methods often fail to emphasize the significance of training strategies like gradient detach, which optimizes the flow of gradients to prevent overfitting the context information.

Gradient detach plays a critical role by blocking the gradients from the context sub-network during the backpropagation process. This ensures that each branch of the network learns distinct and non-overlapping features, enhancing the model's ability to generalize across domains.

Empirical Validation

The paper provides extensive empirical validation that includes ablation studies elucidating the impact of each component in their framework. Results indicate that strategic placement and combination of complementary losses significantly boost performance. Moreover, the proposed method's robustness is evidenced by its applicability across varied and challenging domain adaptation scenarios, including adaptations from synthetic to real-world data.

Among the datasets used, experiments on Cityscapes to FoggyCityscapes illustrate substantial improvements, emphasizing the method's effectiveness in addressing environmental variations like fog in urban scenes. Furthermore, adaptations between dissimilar domains (e.g., PASCAL VOC to Clipart) validate the method's ability to accommodate significant style differences between source and target images.

Implications and Future Directions

The paper's contributions highlight the potential for versatile domain adaptation techniques in real-world object detection applications such as autonomous driving and video surveillance. By ensuring that detection models can adapt to new environments without re-annotations, this approach reduces dependency on manual data labeling and enhances model scalability.

The insights from this research may spur further exploration of training strategies and loss function designs tailored for domain adaptability. Future work could investigate the application of similar strategies for multi-domain or cross-modality adaptations, potentially extending beyond the current focus on visual domains.

Overall, this paper offers a substantial contribution to the field of domain adaptive object detection, providing a pathway toward developing more robust and adaptive machine learning models.