[论文解读] Decoupled Adaptation for Cross-Domain Object Detection

Junguang Jiang, Baixu Chen+2 more|arXiv (Cornell University)|Oct 6, 2021

Advanced Neural Network Applications参考文献 53被引用 28

一句话总结

本文提出 D-adapt，一种用于跨域目标检测的解耦框架，将类别自适应与边界框回归分离，并包含一个边界框适配器以提升定位，在多个跨域任务上取得了最先进的结果。

ABSTRACT

Cross-domain object detection is more challenging than object classification since multiple objects exist in an image and the location of each object is unknown in the unlabeled target domain. As a result, when we adapt features of different objects to enhance the transferability of the detector, the features of the foreground and the background are easy to be confused, which may hurt the discriminability of the detector. Besides, previous methods focused on category adaptation but ignored another important part for object detection, i.e., the adaptation on bounding box regression. To this end, we propose D-adapt, namely Decoupled Adaptation, to decouple the adversarial adaptation and the training of the detector. Besides, we fill the blank of regression domain adaptation in object detection by introducing a bounding box adaptor. Experiments show that D-adapt achieves state-of-the-art results on four cross-domain object detection tasks and yields 17% and 21% relative improvement on benchmark datasets Clipart1k and Comic2k in particular.

研究动机与目标

• Motivate and address domain shift in object detection where multiple objects per image and unknown target labels complicate transfer.
• Decouple adversarial adaptation from detector training to preserve discriminability while enabling transfer.
• Introduce a category adaptor and a bounding box adaptor to separately adapt classification and localization.
• Propose data-distribution adjustments to guide adaptation for category and localization tasks.
• Demonstrate state-of-the-art performance on four cross-domain detection benchmarks and provide ablations elucidating component contributions.

提出的方法

• Decouple cross-domain adaptation into independent sub-problems with separate adaptors (category and bounding box) that are parameter-independent from the detector.
• Introduce a cascading coordination where adaptors leverage information from previous stages but have independent inputs, losses, and parameters.
• Category adaptor uses class-wise adversarial alignment with a confidence-based weighting to respect low-density separation in proposal space.
• Bounding box adaptor employs a regression-based approach with an adversarial regressor to maximize target-domain IoU disparity, guiding a domain-invariant yet accurate box localization.
• Leverage pseudo labeling from adaptors to train the detector on the target domain in an iterative self-feedback loop, without increasing inference cost.
• Algorithm 1 outlines alternating training between proposal generation, adaptor training, and detector optimization; involves Stage 1 source pre-training, Stage 2 category adaptation, Stage 3 box adaptation, and Stage 4 target-domain pseudo-label training.

实验结果

研究问题

• RQ1 Can decoupling category and bounding-box adaptation improve transferability and discriminability in cross-domain object detection?
• RQ2 Does separating adaptation from detector training preserve localization and classification performance under domain shift?
• RQ3 Can a bounding box adaptor improve localization in cross-domain settings beyond category alignment?
• RQ4 How do data-distribution adjustments and pseudo-labeling from adaptors affect target-domain performance?

主要发现

• D-adapt achieves state-of-the-art results on four cross-domain object detection tasks.
• Significant improvements over prior methods on challenging transfers such as VOC-to-Clipart and VOC-to-Comic2k (e.g., substantial mAP gains in reported experiments).
• The category adaptor with a confidence-weighted adversarial objective improves target-domain category discrimination while mitigating ambiguity from background proposals.
• The bounding box adaptor with an IoU-focused disparity objective improves localization, and its adversarial component helps align regression across domains.
• The decoupled framework yields improvements that persist across different detector backbones and settings, with no inference-time overhead since adaptors are detachable.