Domain Adaptive Object Detection via Asymmetric Tri-Way Faster-RCNN

He, Zhenwei; Zhang, Lei

doi:10.1007/978-3-030-58586-0_19

Cited by 112 publications

(60 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Under the KITTI → Cityscape setting, our method achieves the best performance, in terms of both the overall mAP and the performance gain. When adapting from Cityscapes to KITTI, we observe that the results for the "No DA" reported in all the other comparison methods [11], [12], [14], [69] are the same. To reach a fair comparison, we directly report the comparisons on the overall mAP for each method, where our DDF also achieves competitive performance.…”

Section: B Comparison Experimentsmentioning

confidence: 80%

“…Typically, UDA methods transfer the knowledge in four ways: 1) directly minimizing the statistical distribution distance between two domains [22], [23]; 2) inducing the domain-invariant feature generation [9], [10], [24]; 3) learning from the synthesized images [25]- [29], and 4) selftraining via pseudo labels [30], [31]. By alleviating the crossdomain discrepancy at the feature and appearance levels, UDA methods have achieved outstanding performance in crossdomain classification [9], [10], [27], segmentation [32]- [34], and detection [11], [13], [14], [30], [35], [36]. Since our proposed method aims at tackling the feature entanglement issue for cross-domain object detection, only the literature of UDA object detection and feature disentanglement are reviewed in detail.…”

Section: Related Work a Unsupervised Domain Adaptationmentioning

confidence: 99%

“…and F t sha. is induced by optimizing an adversarial domain discriminator at the global level [11], [13], [14] through:…”

Section: B Global Triplet Disentanglementmentioning

confidence: 99%

“…KITTI [66] dataset contains 14999 real autonomous driving images obtained from a mid-sized city, with around 80K bounding boxes annotations for object detection study. Following previous works [11], [12], [14], we employed the training set with 7481 images for all experiments.…”

Section: A Dataset Descriptionmentioning

confidence: 99%

“…(Best viewed in color and zoom in) unlabeled target domain [8]- [10]. For the UDA object detection methods, the majority of them incorporate the adversarial domain discriminators with the object detection architectures (e.g Faster-RCNN), to alleviate the cross-domain discrepancy by inducing the domain-invariant features generation [11]- [14].…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Decompose to Adapt: Cross-domain Object Detection via Feature Disentanglement

Liu¹,

Zhang²,

Yang³

et al. 2022

Preprint

View full text Add to dashboard Cite

Recent advances in unsupervised domain adaptation (UDA) techniques have witnessed great success in cross-domain computer vision tasks, enhancing the generalization ability of data-driven deep learning architectures by bridging the domain distribution gaps. For the UDA-based cross-domain object detection methods, the majority of them alleviate the domain bias by inducing the domain-invariant feature generation via adversarial learning strategy. However, their domain discriminators have limited classification ability due to the unstable adversarial training process. Therefore, the extracted features induced by them cannot be perfectly domain-invariant and still contain domainprivate factors, bringing obstacles to further alleviate the crossdomain discrepancy. To tackle this issue, we design a Domain Disentanglement Faster-RCNN (DDF) to eliminate the sourcespecific information in the features for detection task learning. Our DDF method facilitates the feature disentanglement at the global and local stages, with a Global Triplet Disentanglement (GTD) module and an Instance Similarity Disentanglement (ISD) module, respectively. By outperforming state-of-the-art methods on four benchmark UDA object detection tasks, our DDF method is demonstrated to be effective with wide applicability.

show abstract

Section: B Comparison Experimentsmentioning

confidence: 80%