Learning Domain Adaptive Object Detection with Probabilistic Teacher
Authors: Meilin Chen, Weijie Chen, Shicai Yang, Jie Song, Xinchao Wang, Lei Zhang, Yunfeng Yan, Donglian Qi, Yueting Zhuang, Di Xie, Shiliang Pu
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To validate our approach, we conduct extensive experiments on multiple benchmarks with four different types of domain shifts... Table 1 shows that our method achieves a new SOTA result in both splits. |
| Researcher Affiliation | Collaboration | Meilin Chen 1 Weijie Chen 1 2 Shicai Yang 2 Jie Song 1 Xinchao Wang 3 Lei Zhang 4 Yunfeng Yan 1 Donglian Qi 5 1 Yueting Zhuang 1 Di Xie 2 Shiliang Pu 2 1Zhejiang University 2Hikvision Research Institute 3National University of Singapore 4Chongqing University 5Hainan Institute of Zhejiang University. |
| Pseudocode | Yes | Algorithm 1 Probabilistic Teacher Input: Source domain DS, Target domain DT 1: while Pretraining do 2: Train source only model θI based on Eq. (3) 3: end while 4: θS θI # Duplicate to student model 5: θT θI # Duplicate to teacher model 6: while Mutual Learning do 7: Calculate LS based on Eq. (3) 8: p P L, t P L θT (DT ) # Pseudo labeling, weak augmentation 9: p P L Scls(p P L, τcls) # Sharpen classification probability distributions 10: t P L Sbbox(t P L, τbbox) # Sharpen localization probability distributions 11: p, t θS(DT ) # Feed forward, strong augmentation 12: Calculate EFL LROI T cls based on Eq. (6) and Eq. (11) # Classification adaptation 13: p P L M(p P L) # Merging operation to sum up all foreground probabilities 14: Calculate EFL LRP N T cls based on Eq. (6) and Eq. (11) # Classification adaptation 15: Calculate EFL LROI T bbox based on Eq. (7) and Eq. (11) # Regression adaptation 16: Calculate EFL LRP N T bbox based on Eq. (7) and Eq. (11) # Regression adaptation 17: LT = LRP N T cls + LROI T cls + LRP N T box + LROI T box 18: Ltotal = LS + λT LT 19: Train the anchor shapes via minimizing LT # Anchor adaptation 20: Train student model θS via minimizing Ltotal 21: Update teacher model via EMA 22: end while |
| Open Source Code | Yes | Code available at https://github.com/hikvision-research/Probabilistic Teacher |
| Open Datasets | Yes | Cityscapes (C) (Cordts et al., 2016) contains 2,975 training images and 500 validation images with pixel-level annotations... Foggy Cityscapes (F) (Sakaridis et al., 2018) is a synthetic dataset rendered from Cityscapes... BDD100k (B) (Yu et al., 2018) is a large-scale dataset consisting of 100k images... Sim10k (S) (Johnson-Roberson et al., 2016) consists of 10k images rendered by a gaming engine... KITTI (K) (Geiger et al., 2012) is collected by an autonomous driving platform... |
| Dataset Splits | Yes | Cityscapes (C) (Cordts et al., 2016) contains 2,975 training images and 500 validation images with pixel-level annotations. The annotations are transformed into bounding boxes for the following experiments. Foggy Cityscapes (F) (Sakaridis et al., 2018)... and then report the evaluation results on the validation set of Foggy Cityscapes. BDD100k (B) (Yu et al., 2018)... including 36,728 training and 5,258 validation images, are used for the following experiments. |
| Hardware Specification | No | The paper mentions training on 'a single GPU' but does not provide specific hardware details such as GPU models (e.g., NVIDIA A100, RTX series), CPU models, or cloud computing instance types. |
| Software Dependencies | No | The paper states 'Our implementation is built upon Detectron2 (Wu et al., 2019)' and uses 'torchvision' modules in the appendix, but it does not specify version numbers for Detectron2, torchvision, or any other critical software dependencies. |
| Experiment Setup | Yes | We use a batch size of 16 for both source and target data on a single GPU and train for 30k iterations with a fixed learning rate of 0.016, including 4k iterations for Pretraining and 26k iterations for Multual Learning. The detector is trained by an SGD optimizer with the momentum of 0.9 and the weight decay of 10^-4. The EMA rate α is set to 0.9996. Without careful tuning, the loss weights in this paper are all set to 1. Moreover, λ in EFL, together with temperature τcls and τbbox, are all set to 0.5 simply. |