light_head_rcnn

Examples of Light-Head R-CNN for Object Detection [1]

Performance

MS COCO 2014 minival

Model	mAP@[0.5:0.95] (Original)	mAP@[0.5:0.95] (ChainerCV)
Light-Head R-CNN ResNet101	39.6 % [1] / 40.0 % [2]	39.3 %

Demo

Detect objects in an given image. This demo downloads MS COCO pretrained model automatically if a pretrained model path is not given.

$ python demo.py [--gpu <gpu>] [--pretrained-model <model_path>] <image>.jpg

Evaluation

The evaluation can be conducted using chainercv/examples/detection/eval_coco.py.

Train

You can train the model with the following code. Note that this code requires chainermn module.

$ mpiexec -n <#gpu> python train_multi.py [--batch-size <batch_size>]

References

1. Zeming Li et al. "Light-Head R-CNN: In Defense of Two-Stage Object Detector" ArXiv 2017
2. Light-Head R-CNN