MS-CornerNet

This is a pytorch implementation of MS-CornerNet, a Multispectral extension of the CornerNet paper(https://arxiv.org/abs/1808.01244) to work on RGB+T (Thermal) inputs, specifically the kaist dataset. This repository is extended from the base code at: https://github.com/zzzxxxttt/pytorch_simple_CornerNet and the matlab testing code is taken from https://github.com/Li-Chengyang/MSDS-RCNN with slight modifications.

Requirements:

• python>=3.5
• pytorch==0.4.1 or 1.1.0(DistributedDataParallel training only available using 1.1.0)
• tensorboardX(optional)

Getting Started

1. Disable cudnn batch normalization. Open torch/nn/functional.py and find the line with torch.batch_norm and replace the torch.backends.cudnn.enabled with False.

2. Clone this repo:

3. Compile corner pooling. If you are using pytorch 0.4.1, rename $MS_CORNERNET_ROOT/lib/cpool_old to $MS_CORNERNET_ROOT/lib/cpool, otherwise rename $MS_CORNERNET_ROOT/lib/cpool_new to $MS_CORNERNET_ROOT/lib/cpool.

   cd $CornerNet_ROOT/lib/cpool
   python setup.py install --user
   

4. Compile NMS.

   cd $MS_CORNERNET_ROOT/lib/nms
   make
   

5. For KAIST training, Download KAIST dataset and put data into $CornerNet_ROOT/data/kaist/images and $CornerNet_ROOT/data/kaist/annotations. Annotations should then be further separated into two directories train_sanitized and test_improved

Train

KAIST

multi GPU using nn.parallel.DistributedDataParallel

python -m torch.distributed.launch --nproc_per_node NUM_GPUS train.py \
        --log_name kaist_hg \
        --dataset kaist \
        --arch large_hourglass \
        --lr 5e-4 \
        --lr_step 90,120 \
        --batch_size 8 \
        --num_epochs 100 \
        --num_workers 1

Evaluate

COCO

python test.py --log_name kaist_hg \
               --dataset kaist \
               --arch large_hourglass