Distilling Object Detectors with Feature Richness
Authors: Du Zhixing, Rui Zhang, Ming Chang, xishan zhang, Shaoli Liu, Tianshi Chen, Yunji Chen
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that our methods achieve excellent performance on both anchor-based and anchor-free detectors. For example, Retina Net with Res Net-50 achieves 39.7% in m AP on the COCO2017 dataset, which even surpasses the Res Net-101 based teacher detector 38.9% by 0.8%. |
| Researcher Affiliation | Collaboration | Zhixing Du 1,2,3 Rui Zhang 2,3 Ming Chang 3 Xishan Zhang 2,3 Shaoli Liu 3 Tianshi Chen 3 Yunji Chen 2,4 1University of Science and Technology of China 2SKL of Computer Architecture, Institute of Computing Technology, CAS, Beijing, China 3Cambricon Technologies, China 4University of Chinese Academy of Sciences, China dzx1@mail.ustc.edu.cn {zhangrui,zhangxishan,cyj}@ict.ac.cn {changming,liushaoli,tchen}@cambricon.com |
| Pseudocode | No | The paper describes its methodology using equations and textual explanations, but it does not include any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our implementation is available at https://github.com/duzhixing/FRS. |
| Open Datasets | Yes | We choose the default 120k train images split for training and 5k val images split for the test. Meanwhile, we consider Average Precision as evaluation metric, i.e., m AP, AP50, AP75, APS, APMand APL. (on the COCO dataset [19]) |
| Dataset Splits | No | We choose the default 120k train images split for training and 5k val images split for the test. The paper specifies train and test splits, but a distinct validation split for model tuning during training is not explicitly mentioned. |
| Hardware Specification | No | The paper discusses model deployment on 'resource-limited devices' and mentions 'large-scale deep models' but does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for running the experiments. |
| Software Dependencies | No | Our implementation is based on mmdetection [4] with Pytorch framework. While it names the frameworks used, specific version numbers for mmdetection or Pytorch are not provided. |
| Experiment Setup | Yes | we use 2x learning schedule to train 24 epochs or the 1x learning schedule to train 12 epochs on COCO dataset. And the learning rate is divided by 10 at the 8-th and 11-th epochs for 1x schedule and the 16-th and 22-th epochs for 2x schedule. We set momentum as 0.9 and weight decay as 0.0001. |