detectron/projects/TridentNet

• TridentNet in Detectron2
• • Training
  • Evaluation
  • Results on MS-COCO in Detectron2
  • Citing TridentNet

TridentNet in Detectron2

Scale-Aware Trident Networks for Object Detection

Yanghao Li*, Yuntao Chen*, Naiyan Wang, Zhaoxiang Zhang

[TridentNet] [arXiv] [BibTeX]

In this repository, we implement TridentNet-Fast in Detectron2.
Trident Network (TridentNet) aims to generate scale-specific feature maps with a uniform representational power. We construct a parallel multi-branch architecture in which each branch shares the same transformation parameters but with different receptive fields. TridentNet-Fast is a fast approximation version of TridentNet that could achieve significant improvements without any additional parameters and computational cost.

Training

To train a model, run

python /path/to/detectron2/projects/TridentNet/train_net.py --config-file <config.yaml>

For example, to launch end-to-end TridentNet training with ResNet-50 backbone on 8 GPUs,
one should execute:

python /path/to/detectron2/projects/TridentNet/train_net.py --config-file configs/tridentnet_fast_R_50_C4_1x.yaml --num-gpus 8

Evaluation

Model evaluation can be done similarly:

python /path/to/detectron2/projects/TridentNet/train_net.py --config-file configs/tridentnet_fast_R_50_C4_1x.yaml --eval-only MODEL.WEIGHTS model.pth

Results on MS-COCO in Detectron2

Model	Backbone	Head	lr sched	AP	AP50	AP75	APs	APm	APl	download
Faster	R50-C4	C5-512ROI	1X	35.7	56.1	38.0	19.2	40.9	48.7	model | metrics
TridentFast	R50-C4	C5-128ROI	1X	38.0	58.1	40.8	19.5	42.2	54.6	model | metrics
Faster	R50-C4	C5-512ROI	3X	38.4	58.7	41.3	20.7	42.7	53.1	model | metrics
TridentFast	R50-C4	C5-128ROI	3X	40.6	60.8	43.6	23.4	44.7	57.1	model | metrics
Faster	R101-C4	C5-512ROI	3X	41.1	61.4	44.0	22.2	45.5	55.9	model | metrics
TridentFast	R101-C4	C5-128ROI	3X	43.6	63.4	47.0	24.3	47.8	60.0	model | metrics

Citing TridentNet

If you use TridentNet, please use the following BibTeX entry.

@InProceedings{li2019scale,
  title={Scale-Aware Trident Networks for Object Detection},
  author={Li, Yanghao and Chen, Yuntao and Wang, Naiyan and Zhang, Zhaoxiang},
  journal={The International Conference on Computer Vision (ICCV)},
  year={2019}
}