EAI_self_driving

Faster rcnn

Meta Faster rcnn

Data

BDD Dataset - Paper

Google Drive

Pretrain model

1. For vgg16 Here
2. For RegNet101 Here

Before getting start

1. Create data folder
2. Create pretrained_model folder and put Pretrain model in it
3. Create VOCdevkit2007 folder and put VOC2007 in it

data
  |----- pretrained_model
          |----- vgg16_caffe.pth
          |----- resnet101_caffe.pth
  |----- VOCdevkit2007
          |----- VOC2007

Install

pip install -r requirements.txt
cd lib
sh make.sh (linux / mac)     |  python setup.py build develop (windows)

Training on pascal_voc

CUDA_VISIBLE_DEVICES=0 python trainval_net.py \
               --dataset pascal_voc --net res101 \
               --epochs 1 --bs 4 --nw 0 \
               --lr 1e-3 --lr_decay_step 5 \
               --cuda

Keep training

CUDA_VISIBLE_DEVICES=0 python trainval_net.py \
               --epochs 6 --cuda \
               --r True --checksession 1 --checkepoch 4 --checkpoint 79855               

Test (mAP)

python test_net.py --dataset pascal_voc --net res101 \
               --checksession 1 --checkepoch 4 --checkpoint 79855 \
               --cuda

Result (Image)

python demo.py --net res101 \
           --checksession 1 --checkepoch 4 --checkpoint 79855 \
           --cuda --load_dir models