Decoupled Adaptation for Cross-Domain Object Detection

Authors: Junguang Jiang, Baixu Chen, Jianmin Wang, Mingsheng Long

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that D-adapt achieves state-of-the-art results on four crossdomain object detection tasks and yields 17% and 21% relative improvement on benchmark datasets Clipart1k and Comic2k in particular.
Researcher Affiliation Academia Junguang Jiang, Baixu Chen, Jianmin Wang, Mingsheng Long School of Software, BNRist, Tsinghua University, China {jjg20,chenbx18}@mails.tsinghua.edu.cn, {jimwang,mingsheng}@tsinghua.edu.cn
Pseudocode Yes Algorithm 1: D-adapt Training Pipeline.
Open Source Code Yes Code is available at https://github. com/thuml/Decoupled-Adaptation-for-Cross-Domain-Object-Detection.
Open Datasets Yes Following six object detection datasets are used: Pascal VOC [11], Clipart [21], Comic [21], Sim10k [23], Cityscapes [9] and Foggy Cityscapes [44].
Dataset Splits Yes Comic2k contains 1k training images and 1k test images... Both Cityscapes and Foggy Cityscapes have 2975 training images and 500 validation images with 8 object categories.
Hardware Specification Yes We perform all experiments on public datasets using a 1080Ti GPU.
Software Dependencies No The paper mentions frameworks and models like Faster-RCNN, Res Net101, VGG-16, and SGD optimizer, but does not provide specific version numbers for any software libraries or dependencies (e.g., PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes Stage 1: Source-domain pre-training... with a learning rate of 0.005 for 12k iterations. Stage 2: Category adaptation... trained for 10k iterations using SGD optimizer with an initial learning rate of 0.01, momentum 0.9, and a batch size of 32 for each domain... λ is kept 1 for all experiments. Stage 3: Bounding box adaptation... The training hyper-parameters (learning rate, batch size, etc.) are the same as that of the category adaptor. η is kept 0.1 for all experiments. Stage 4: Target-domain pseudo-label training... for 4k iterations, with an initial learning rate of 2.5 10 4 and reducing to 2.5 10 5 exponentially. The adaptors and the detector are trained in an alternative way for T = 3 iterations.