Training-Time-Friendly Network for Real-Time Object Detection
Authors: Zili Liu, Tu Zheng, Guodong Xu, Zheng Yang, Haifeng Liu, Deng Cai11685-11692
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on MS COCO show that our TTFNet has great advantages in balancing training time, inference speed, and accuracy. |
| Researcher Affiliation | Collaboration | 1State Key Lab of CAD&CG, Zhejiang University, Hangzhou, China 2Fabu Inc., Hangzhou, China 3Alibaba-Zhejiang University Joint Institute of Frontier Technologies, Hangzhou, China |
| Pseudocode | No | The paper describes its methods and architecture but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code has been made available at https://github.com/ZJULearning/ttfnet. |
| Open Datasets | Yes | Our experiments are based on the challenging MS COCO 2017 benchmark. We use the Train split (115K images) for training and report the performances on Val split (5K images). |
| Dataset Splits | Yes | We use the Train split (115K images) for training and report the performances on Val split (5K images). |
| Hardware Specification | Yes | Our experiments are based on open source detection toolbox mmdetection (Chen et al. 2019) with 8 GTX 1080Ti. |
| Software Dependencies | No | The paper mentions using the 'mmdetection' toolbox but does not provide specific version numbers for it or any other software dependencies like deep learning frameworks or libraries. |
| Experiment Setup | Yes | We resize the images to 512 512 and do not keep the aspect ratio. Only the random flip is used for data augmentation in training unless the long training schedule is adopted. We use unfrozen BN but freeze all parameters of stem and stage 1 in the backbone. For Res Net, the initial learning rate is 0.016, and the mini-batch size is 128. For Dark Net, the initial learning rate is 0.015, and the mini-batch size is 96. The learning rate is reduced by a factor of 10 at epoch 18 and 22, respectively. Our network is trained with SGD for 24 epochs. For the super-fast version, the training schedule is halved. For the long-training version, the training schedule is increased by five times(i.e., 120-epochs training in total). Weight decay and momentum are set as 0.0004 and 0.9, respectively. For bias parameters in the network, their weight decay is set to 0, and their learning rate is doubled. Warm-up is applied for the first 500 steps. |