Torchvision Faster R-CNN Pre-trained on the COCO dataset

This repository aims to showcase a model of the Faster RCNN detector[1] pre-trained on the COCO dataset[2]. The implementation is the one in vision. The training was done using the scripts from the detection folder in the vision repository. The model was trained in a rig with 4 GPUs. The only change to the default parameters was the number of images per GPU, that was set to 4 to match a batch size of 16 images.

The pretrianed model is hosted here.

Performance

IoU metric: bbox

Metric	IoU	area	maxDets	Result
Average Precision (AP)	0.50:0.95	all	100	0.356
Average Precision (AP)	0.50	all	100	0.560
Average Precision (AP)	0.75	all	100	0.384
Average Precision (AP)	0.50:0.95	small	100	0.178
Average Precision (AP)	0.50:0.95	medium	100	0.389
Average Precision (AP)	0.50:0.95	large	100	0.486
Average Recall (AR)	0.50:0.95	all	1	0.302
Average Recall (AR)	0.50:0.95	all	10	0.459
Average Recall (AR)	0.50:0.95	all	100	0.477
Average Recall (AR)	0.50:0.95	small	100	0.257
Average Recall (AR)	0.50:0.95	medium	100	0.511
Average Recall (AR)	0.50:0.95	large	100	0.646

Example

To see how to load the model and compute the performance metrics, look at the evaluate.ipynb notebook

References

[1] Ren, Shaoqing, et al. "Faster r-cnn: Towards real-time object detection with region proposal networks." Advances in neural information processing systems. 2015.

[2] Lin, Tsung-Yi, et al. "Microsoft coco: Common objects in context." European conference on computer vision. Springer, Cham, 2014.