Marine Animal Detection

This repository is built for UCSD ECE 228 final project, maintained by Di Gu, Shuangcheng Yang, Yixun Zhang and Kunmao Li. Special thanks to OBSEA for sharing their underwater camera image with us.
Please refer this paper for more details on our project.

File Organization

• FasterRCNN - a pytorch version of Faster RCNN
• Results - detection results from models
• YOLOv3 - a pytorch version of YOLOv3
• data - training dataset with cfg files for YOLOv3
• tools - codes for data processing, visualization

Model 1: PyTorch-YOLOv3

A minimal PyTorch implementation of YOLOv3, with support for training, inference and evaluation. This file is a reversion of yolov3 and most of the py files are coming from another github author. We test and try to refrain the work. If you want to know more, please go to this repo.

Installation

Clone and install requirements

$ git clone https://github.com/Di-Gu/Detection.git
$ cd Detection/YOLOv3
$ sudo pip3 install -r requirements.txt

Except these three commend lines, user also need to use the newest vision of torchvision. 
The easiest way is delete the original torchvision in datahub. Then use the following commend 
to set up the newest vision of tochvision. The git hub link is showing as following.
https://github.com/pytorch/vision.git

one can install this easily.
$ pip install torchvision

Download pretrained weights (Run this only if you want to train dataset. For test demo, user don't need to type these commands)

$ cd weights/
$ bash download_weights.sh

Train

Google colab is suggested to run the 'ece228project.ipynb' file. After you finish set up environment, you can start to training. You can find the dataset in data folder. After setting up the dataset loading path, just run all the files and you can see the training table show up. It may take hours to finish all the epochs.

Test Demo

After training, you can use the saved '.pth' weight file to run the whole model to see the test result in the same ipynb file.

Model 2: Faster RCNN

The pytorch implementation of Faster RCNN trained on own dataset, The original work is obtained from Jianwei Yang.

Installation

Find pre-trained weights on here, if you want to train from begining.

Prerequisites

Pytorch 1.0
CUDA 10.1

Compilation

$ cd FasterRCNN
$ pip install -r requirements 
$ cd lib
$ python setup.py build develop

Train

For this project, trainging is done on Manjaro Linux with GTX 1080 Ti

After setup the environment, start training with:

$ CUDA_VISIBLE_DEVICES=0 python trainval_net.py --dataset pascal_voc --net res101 --bs 4 --lr 4e-3 --lr_decay_step 8 --cuda

Test Demo

Download trained weights here. Then run following line for testing

$ python demo.py --checksession 1 --checkepoch 20 --checkpoint 139 --cuda --load_dir models

Model 3: C++ based YOLOv3

A rework of the most detailed implementation of YOLOv3 obtained from AlexeyAB.

Installation

Following the step from the repo
After successfully compiled on Windows
Download the pretrained weight for the convolutional layers from here
Relocate the cfg files in the data folder accordingly

Train

On Windows powershell run

.\darknet.exe detector train data/obj.data yolo-obj.cfg darknet53.conv.74 -map 2>1 | tee train_yolov3.log

Test Demo

Download trained weights here. Then run following line for testing

darknet.exe detector test data/obj.data yolo-obj-test.cfg backup/yolo-obj_last.weights data/test/test_3.jpg -thresh 0.5