IDRNet: Intervention-Driven Relation Network for Semantic Segmentation

Authors: Zhenchao Jin, Xiaowei Hu, Lingting Zhu, Luchuan Song, Li Yuan, Lequan Yu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments are conducted to validate the effectiveness of IDRNet quantitatively and qualitatively.
Researcher Affiliation Collaboration 1The University of Hong Kong 2Shanghai AI Laboratory 3Peking University 4University of Rochester
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks. Figure 2 is a diagram of the network.
Open Source Code Yes Code is available at https://github.com/SegmentationBLWX/sssegmentation.
Open Datasets Yes Our approach is validated on five popular semantic segmentation benchmark datasets, including ADE20K [87], COCO-Stuff [4], Cityscapes [17], LIP [29] and PASCAL-Context [22].
Dataset Splits Yes In detail, ADE20K is one of the most well-known datasets for scene parsing, which contains 150 stuff/object category labels. There are 20K/2K/3K images for training, validation and test set, respectively in the dataset. COCO-Stuff...It consists of 9K/1K images in the training and test sets. Cityscapes has 5K high-resolution annotated urban scene images, with 2,975/500/1,524 images for training/validation/testing. LIP mainly focuses on single human parsing and contains 50K images with 19 semantic human part classes and 1 background class. Its training, validation and test sets separately involve 30K/10K/10K images. PASCAL-Context...The dataset is divided into 4,998 and 5,105 images for training and validation.
Hardware Specification Yes Table 5: Complexity comparison with existing context modules on a single RTX 3090 Ti GPU.
Software Dependencies No Our algorithm is implemented in Py Torch [54] and SSSegmentation [38]. No specific version numbers for PyTorch or SSSegmentation are provided.
Experiment Setup Yes Specific to ADE20K, we set learning rate, crop size, batch size and training epochs as 0.01, 512 × 512, 16 and 130, respectively. Specific to COCO-Stuff, learning rate, crop size, batch size and training epochs are set as 0.001, 512 × 512, 16 and 110, respectively. As for LIP, we set learning rate, crop size, batch size and training epochs as 0.01, 473 × 473, 32 and 150, respectively. As for Cityscapes, learning rate, crop size, batch size and training epochs are set as 0.01, 512 × 1024, 8 and 220, respectively. Specific to PASCAL-Context, we set learning rate, crop size, batch size and training epochs as 0.004, 480 × 480, 16 and 260, respectively.