FreeAnchor: Learning to Match Anchors for Visual Object Detection
Authors: Xiaosong Zhang, Fang Wan, Chang Liu, Rongrong Ji, Qixiang Ye
NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on COCO demonstrate that Free Anchor consistently outperforms the counterparts with significant margins. Detectors were trained on COCO training set, and evaluated on the val set. Final results were reported on the test-dev set. Table 2: Performance comparison of Free Anchor and Retina Net (baseline). |
| Researcher Affiliation | Academia | 1University of Chinese Academy of Sciences, Beijing, China 2Xiamen University, Xiamen, China 3Peng Cheng Laboratory, Shenzhen, China |
| Pseudocode | Yes | Algorithm 1 Detector training with Free Anchor. |
| Open Source Code | Yes | 1Code is available at https://github.com/zhangxiaosong18/Free Anchor |
| Open Datasets | Yes | Experiments were carried out on COCO 2017[19], which contains 118k images for training, 5k for validation (val) and 20k for testing without provided annotations (test-dev). |
| Dataset Splits | Yes | Experiments were carried out on COCO 2017[19], which contains 118k images for training, 5k for validation (val) and 20k for testing without provided annotations (test-dev). |
| Hardware Specification | Yes | Training used synchronized SGD over 8 Tesla V100 GPUs with a total of 16 images per mini-batch (2 images per GPU). |
| Software Dependencies | No | The paper mentions implementing Free Anchor upon Retina Net [7] and using Res Net [20] and Res Ne Xt [21] as backbones, but does not specify software versions for these or other dependencies like Python, PyTorch, or TensorFlow. |
| Experiment Setup | Yes | For the last convolutional layer of the classification subnet, we set the bias initialization to b = log ((1 ρ)/ρ) with ρ = 0.02. Unless otherwise specified, all models were trained for 90k iterations with an initial learning rate of 0.01, which is then divided by 10 at 60k and again at 80k iterations. Anchor bag size n: We evaluated anchor bag sizes in {40, 50, 60, 100} and observed that the bag size 50 reported the best performance. Background Io U threshold t: A threshold was used in P{aj bi} during training. We tried background Io U thresholds in {0.5, 0.6, 0.7} and validated that 0.6 worked best. Focal loss parameter: ...we set α = 0.5 and γ = 2.0. Loss regularization factor β: The regularization factor β in Eq. 1, which balances the loss of classification and localization, was experimentally validated to be 0.75. |