Fine-Grained Dynamic Head for Object Detection
Authors: Lin Song, Yanwei Li, Zhengkai Jiang, Zeming Li, Hongbin Sun, Jian Sun, Nanning Zheng
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the effectiveness and efficiency of the proposed method on several state-of-the-art detection benchmarks. Code is available at https://github.com/Steven Grove/Dynamic Head. Overall, the proposed dynamic head is fundamentally different from existing methods for head designs. Our approach exploits a new dimension: dynamic routing mechanism is utilized for fine-grained object representation with efficiency. The designed method can be easily instantiated on several FPN-based object detectors [2, 3, 5, 20, 21] for better performance. Moreover, extensive ablation studies have been conducted to elaborate on its superiority in both effectiveness and efficiency, which achieve consistent improvements with little computational overhead. For instance, with the proposed dynamic head, the FCOS [3] based on the Res Net-50 [28] backbone attains 2.3% m AP absolute gains with less computational cost on the COCO [29] dataset. |
| Researcher Affiliation | Collaboration | 1 College of Artificial Intelligence, Xi an Jiaotong University 2 The Chinese University of Hong Kong 3 Institute of Automation, Chinese Academy of Sciences 4 Megvii Inc. (Face++) |
| Pseudocode | No | The paper does not include pseudocode or a clearly labeled algorithm block. |
| Open Source Code | Yes | Code is available at https://github.com/Steven Grove/Dynamic Head. |
| Open Datasets | Yes | All the backbones are pre-trained on the Image Net classification dataset [38]. All the experiments are trained on 8 GPUs with 2 images per GPU (effective mini-batch size of 16) for 90K iterations. All the backbones are pre-trained on the Image Net classification dataset [38]. All the training hyper-parameters are identical to the 1x schedule in the Detectron2 [40] framework. All the models are optimized by using Synchronized SGD [41] with a weight decay of 0.0001 and a momentum of 0.9. All the reported results here are based on Res Net-50 [28] backbone and evaluated on COCO val set. |
| Dataset Splits | Yes | All the backbones are pre-trained on the Image Net classification dataset [38]. All the training hyper-parameters are identical to the 1x schedule in the Detectron2 [40] framework. All the reported results here are based on Res Net-50 [28] backbone and evaluated on COCO val set. |
| Hardware Specification | No | The paper mentions that "All the experiments are trained on 8 GPUs" but does not specify the model or type of GPUs, CPUs, or any other specific hardware components. |
| Software Dependencies | No | The paper mentions "Detectron2 [40] framework" but does not specify its version or any other software dependencies with version numbers. |
| Experiment Setup | Yes | All the backbones are pre-trained on the Image Net classification dataset [38]. Batch normalizations [39] in the backbone are frozen. In the training phase, input images are resized so that the shorter side is 800 pixels. All the training hyper-parameters are identical to the 1x schedule in the Detectron2 [40] framework. Specifically, we fix parameters of the first two stages in the backbone and then jointly finetune the rest network. All the experiments are trained on 8 GPUs with 2 images per GPU (effective mini-batch size of 16) for 90K iterations. The learning rate is initially set to 0.01 and then decreased by 10 at the 60K and 80K iterations. All the models are optimized by using Synchronized SGD [41] with a weight decay of 0.0001 and a momentum of 0.9. |