Co-mining: Self-Supervised Learning for Sparsely Annotated Object Detection
Authors: Tiancai Wang, Tong Yang, Jiale Cao, Xiangyu Zhang2800-2808
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments are performed on MS COCO dataset with three different sparsely annotated settings using two typical frameworks: anchorbased detector Retina Net and anchor-free detector FCOS. Experimental results show that our Co-mining with Retina Net achieves 1.4 % 2.1% improvements compared with different baselines and surpasses existing methods under the same sparsely annotated setting. |
| Researcher Affiliation | Collaboration | Tiancai Wang1, Tong Yang1, Jiale Cao2, Xiangyu Zhang1 1 MEGVII Technology 2 Tianjin University {wangtiancai, yangtong, zhangxiangyu}@megvii.com, connor@tju.edu.cn |
| Pseudocode | Yes | Algorithm 1: Our Co-mining Algorithm |
| Open Source Code | No | The paper does not provide an explicit statement or link for the open-source code of the Co-mining methodology. It only mentions re-implementing other methods due to lack of their source code. |
| Open Datasets | Yes | Experiments are conducted on the sparsely annotated versions of MS COCO dataset (Lin et al. 2014). The COCO-2017 train set with sparse annotations is used for training. |
| Dataset Splits | Yes | The COCO-2017 train set with sparse annotations is used for training. The COCO-2017 validation set with complete annotations is used for all performance evaluations. |
| Hardware Specification | Yes | We adopt 8 TITAN 2080ti GPUs with a batch size of 16 for training. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies, such as deep learning frameworks or programming languages. |
| Experiment Setup | Yes | During training, there are 90k iterations in total. The learning rate is initially set to 0.01 and gradually decreases to 0.001 and 0.0001 at 60k and 80k iterations. Warm-up strategy adopted for the first 1k iterations to stabilize the training process. |