This is the reimplementation of the Faster RCNN based on Progress Lab Dataset, the original one could be referred to official code. I cleaned the redundant code for specific usage of 4Progress Lab.

There are two usages of this program, one is used for training end to end faster rcnn model, the other one is used for testing the model trained. Also with the following instruction, you could train your own dataset.

1. Constuct your dataset structure and follow the dataset structure like VOC2012, and put it in /data folder like /progressiros folder in there
2. In /data/scripts folder, run fetch_imagenet_models.sh to fetch pretrained imagenet model and put it into /data/faster_rcnn_models
3. In /lib/models/progressiros/VGG16/faster_rcnn_end2end, modify two files named test.protxt and train.protxt to specify your dataset. In test.protxt file, modify cls_score and bbox_pred layers output to number of your dataset classes and 4*(num of your dataset classes)
4. In /lib/datasets/tools folder, in progress.py, modify name of classes
5. specify pretrained model in train_progress.sh

Then run following command in ROOT FOLDER to begin to train

sh train_progress.sh

Attention!!! Before that!!!

Please include caffe module in ROOT FOLDER caffe-fast-rcnn and compile that!!!

The trained model will be in /output/faster_rcnn_end2end/train folder

1. Please put test image in /data/demo folder
2. In /tools/ folder, in demo.py file, modify class name and model name you want to use
3. Run demo.py file to see the results and specify pretrianed model

1. specify pretrained model in test_progress.sh
2. Run sh test_progress.h
3. The default setting for calculating mAP is to read results from text file and compare them to xml file, you can change the setting in experiments/scripts/faster_rcnn_end2end.sh by setting --txt to 0. Then it well detect the image and calculate the mAP
4. The format of results of detection is like following:

obj1 xmin ymin xmax ymax
obj2 xmin ymin xmax ymax
...

1. You may have to change path in lib/datasets/tools/factory.py for training
2. You may have to change path in line 258 in lib/fast_rcnn/test.py and line 85 in tools/train_net.py for reading img and detection results path for testing

For more detail, please email ywchow@umich.edu

More features would be released soon!