SCAN: Cross Domain Object Detection with Semantic Conditioned Adaptation
Authors: Wuyang Li, Xinyu Liu, Xiwen Yao, Yixuan Yuan1421-1428
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Comprehensive experiments conducted on three adaptation benchmarks demonstrate that SCAN outperforms existing works by a large margin. We conduct extensive experiments on three domain adaptation scenarios, following the standard settings in literature (Hsu et al. 2020a), i.e. training with labeled source data and unlabeled target data, and testing on the target data. |
| Researcher Affiliation | Academia | Wuyang Li1, Xinyu Liu1, Xiwen Yao2, Yixuan Yuan1* 1 City University of Hong Kong 2 Northwestern Polytechnical University |
| Pseudocode | No | The paper describes methods in text and provides equations, but does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Open source: https://github.com/City U-AIM-Group/SCAN. |
| Open Datasets | Yes | Cityscapes (Cordts et al. 2016) is a city landscape dataset under dry weather condition with eight annotated categories, which consists train set with 2975 images and validation set with 500 images. Foggy Cityscapes (Sakaridis, Dai, and Van Gool 2018) is a synthesized dataset from Cityscapes as foggy weather. Sim10k (Johnson-Roberson et al. 2017) is a simulated dataset with 10,000 images with the labels of annotated car bounding boxes. KITTI (Geiger, Lenz, and Urtasun 2012) is a real-world scene dataset collected with different camera setups. |
| Dataset Splits | Yes | Cityscapes (Cordts et al. 2016) is a city landscape dataset under dry weather condition with eight annotated categories, which consists train set with 2975 images and validation set with 500 images. training with labeled source data and unlabeled target data, and testing on the target data. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper mentions software components and models like VGG16 and FCOS, but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | We adopt the Stochastic Gradient Descent (SGD) optimizer with a 0.0025 learning rate and an 8 batch-size. α and β are set 0.1 and 1, respectively. The nonlinear projection is deployed with the Conv ReLU-Conv structure, and the nonlinear classifier fs is in the Fc-ReLU-Fc format. The three-dimensional paradigm P records T=3 iterations and uses D=256 channels to model class-specific distributions. |