ya_mxdet provides a simple Faster R-CNN (proposed in Faster R-CNN) implementation fully in MXNet gluon API. More functions are in developing.
ya_mxdet is not exactly the re-implementation of Faster R-CNN. You may need to tune it carefully for your tasks.
ya_mxdet depends on MXNet 0.12 with GPU support.
Refer to https://mxnet.incubator.apache.org/get_started/install.html for more details.
OpenCV is also required. Installation from source is highly recommended, since in the repo of some linux distribution OpenCV does not support imshow().
Before you train a model, you need to prepare Pascal VOC 2007 dataset. You can download it from http://host.robots.ox.ac.uk/pascal/VOC/. Then set a soft link VOC2007Train pointing to VOCdevkit/VOC2007.
You also need a GPU with at least 11GB memory. The code may download VGG16 pretrained model at the first run.
$ mkdir model_dump $ python3 train_faster_rcnn.py
You can get detection result by:
$ python3 inference_faster_rcnn.py [model_file]
The results are stored in inference_output.json by default, which is in the format of COCO results format.
voc_ruller.py is used to calculate mAP metric of PASCAL VOC 2007 on a inference output file.
All the code are distributed with a MIT-style license. If you use this repo in your research, it would be nice to cite the original Faster R-CNN paper and add the link of this project into your report.
In following samples, the blue boxes are the ground truth while the red boxes are the detection result.
Only the boxes with scores higher than 0.5 are shown in the demo.
Only 4 boxes after NMS are shown in the demo.
