training-domain-specific-models

Training domain specific models for efficient object detection arXivpaper

This is a framework to train domain specific model, which is accurate + computation efficient!

Faster-RCNN implementation is based on faster-rcnn.pytorch by jwyang. Thanks!

I strongly recommend to take a look at their readme if you get stuck on frcnn codes.

https://github.com/jwyang/faster-rcnn.pytorch

What is a domain specific model?

For a backbone of object detection, Resnet101 is a very good model but TOO BIG! But a small model like Resnet18 has a low accuracy, due to small network capacity.

However, do we need such a big model to do object detection in a limited domain? (Like your office or a particular intersection) Since backgrounds does not change, even small model should do very well if trained properly!

A domain specific model(DSM) is a model focusing on achieving high accuracy at such limited domain (e.g. fixed view of an intersection). We argue that DSMs can capture essential features well even with a small model size.

In this repo, we train a small domain specific model (say res18) in with a dataset of a limited domain.

We see that by training, small models can achieve very high accuracy!

Take a look at a Youtube Video Demo!

By domain specific training, the mAP improves ~20%.

Preparation

Requirements.

Pytorch 0.4.0

Python 3.x

CUDA 8.0 or higher

Clone repo

Lets start off by cloning this repo.

git clone https://github.com/kentaroy47/training-domain-specific-models.git
cd training-domain-specific-models

You may need to compile the rpn scripts.

Please see jwyang's repo for details.

https://github.com/jwyang/faster-rcnn.pytorch

Download models

We need to prepare Resnet101 and Resnet18 Faster-RCNN model.

cd training-domain-specific-models

Download.. 
wget https://www.dropbox.com/s/ew47jhdu67bdocf/files.tar.gz
tar -zxvf files.tar.gz

Setup dataset

If the models and the video are set, we can prepare the dataset.

1. Res101 model generates the teacher labels.
2. The dataset is prepared in a PASCAL_VOC format for training.

This is done in a single script.

Just run:

# for dataset coral
python make_dataset.py　--dataset coral
# for dataset jackson2
python make_dataset.py　--dataset jackson2

shortcut..

We prepared a dataset.tar in the link bellow, if you want to take a short cut.

Actually cloning the repo will get you the pickle lable files (output/baseline/)

Training Domain Specific Models!

Run..

This will take about 2 hours on TitanXp.

python trainval_net_ds.py --cuda --r True --dataset pascal_voc_jackson2

# or for coral,
python trainval_net_ds.py --cuda --r True --dataset pascal_voc_coral

Evaluation!

We evaluate the accuracy (mAP) with validation images.

The res101 outputs are utilized as ground truth here, since labeling them are cubersome.

python demo-and-eval-save.py --net res18 --dataset pascal_voc_jackson2 --cuda --checksession 1 --checkepoch 20 --checkpoint 1 --image_dir images/jackson2_val/ --truth output/baseline/jackson2val-res101.pkl

Ground truths

We also plan to release the hand-labeled ground truth as well.

Interestingly, domain specific model outperforms the accuracy that of res101.