Faster RCNN

To run this code, you need to install mxnet and other dependencies first. To do that, you can run

bash script/additional_deps.sh

to install MXNet and other required packages of python.

Then, you need to run

bash script/get_voc.sh
bash script/get_pretrained_model.sh

to get VOC data and pretrained VGG-16 network.

To download the KITTI dataset you need to request the download link, hence we do not provide a script here. The data should have the structure like

├── kitti
│   ├── images
│   ├── imglists
│   └── results

Our own data is not public, hence we can only present the result. We can not release the dataset.

VOC

To train on the VOC-07 data using pretrained VGG-16, run

python train_end2end.py --image_set 2007_trainval --gpu 0

To use both VOC-07 and VOC-12 data, run

python train_end2end.py --image_set 2007_trainval+2012_trainval --gpu 0

After train completed, run

python test.py --gpu 0

to test on VOC-07 test dataset.

KITTI

To train on the KITTI data using model pretrained on VOC, uncomment

# del arg_params['cls_score_weight'], arg_params['cls_score_bias']
# del arg_params['bbox_pred_weight'], arg_params['bbox_pred_bias']

in load_param function in rcnn/utils/load_model.py

Then, run

python train_end2end.py --dataset Kitti --pretrained e2e --pretrained_epoch 10 --prefix e2e_kitti

After train completed, run

python test.py --dataset Kitti --image_set test --gpu 0 --prefix e2e_kitti --epoch 10 --thresh 0.01

then submit the result to KITTI website to get the result.

Own Dataset

Since the data is not public yet, we do not provide the script to train/test it. However it is very similar to the KITTI dataset.