Augmentation-Free Dense Contrastive Knowledge Distillation for Efficient Semantic Segmentation

Authors: Jiawei Fan, Chao Li, Xiaolong Liu, Meina Song, Anbang Yao

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on five mainstream benchmarks with various teacher-student network pairs demonstrate the effectiveness of our approach. Experimental results demonstrate that: (i) Af-DCD exhibits superior performance compared to state-of-the-art methods, on various benchmarks with different teacher-student network pairs; (ii) Af-DCD exhibits even more significant improvements on larger datasets, such as ADE20K, indicating it can enhance student s generalization capability.
Researcher Affiliation Collaboration Jiawei Fan Intel Labs China jiawei.fan@intel.com Chao Li Intel Labs China chao3.li@intel.com Xiaolong Liu Holo Matic Technology Co. Ltd. liuxiaolong@holomatic.com Meina Song BUPT mnsong@bupt.edu.cn Anbang Yao Intel Labs China anbang.yao@intel.com
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/OSVAI/Af-DCD.
Open Datasets Yes Five popular semantic segmentation datasets, including Cityscapes [27], Pascal VOC [28], Camvid [29], ADE20K [30] and COCO-Stuff-164K [31], are used in our experiments.
Dataset Splits Yes Five popular semantic segmentation datasets, including Cityscapes [27], Pascal VOC [28], Camvid [29], ADE20K [30] and COCO-Stuff-164K [31], are used in our experiments. Following general settings [9, 20, 15] in semantic segmentation distillation...
Hardware Specification Yes The training time is measured on 8 NVIDIA RTX A5000 GPUs with 40000 iterations.
Software Dependencies Yes we implement our method on both MMSegmentation codebase [35] and CIRKD codebase [9].
Experiment Setup Yes In training phase, all models are optimized by SGD with the momentum of 0.9, the initial learning rate of 0.02, and the batch size of 16. The input size is 512 × 1024, 400 × 400, 512 × 1024, 512 × 1024, for experiments on Pascal VOC, Cam Vid, ADE20K and COCO-Stuff-164K, respectively. The input size for experiments on Cityscapes are different in the two codebase, 512 × 1024 in CIRKD codebase and 512 × 512 in MMSegmentation codebase [15]. Our masked reconstruction generator consists of two 3 × 3 convolutional layers with Re LU, following [15].