Object Detection Laboratory

This project is my library of One-stage Object Detection. My code is clean and concise, without too deep encapsulation, so it is easy to understand the function of each module.

For example,

• If you want to know how to pull data from COCO, you just open dataset/coco.py.
• If you want to know how to build RetinaNet, you just open models/detector/retinanet/ and config/retinanet_config.py.
• If you want to know how to build FPN, you just open models/neck/fpn.py/.
• If you want to know the whole pipeline of training, you just open train.py.
• If you want to know the whole pipeline of evaluation, you just open eval.py and evaluator/coco_evaluator.py.
• If you want to know how to visualize the detection results on detection benchmark like COCO or VOC, you just open test.py.
• If you want to know how to run a demo with images or videos on your own device, you just open demo.py.
• If you want to know how to run a demo with your own camero, you just also open demo.py.

So, it is very friendly, right?

I am sure you will soon become familiar with this benchmark and add your own modules to it.

However, limited by my computing resources, I cannot use larger backbone networks like ResNet-101 and ResNeXt-101 to train more powerful detectors. If you have sufficient computing resources and are already using these larger backbone networks to train the detectors in this benchmark, I look forward to your open source weight files to complement this project. Thanks a lot.

Coming soon

• [x] YOLOF
• [x] RetinaNet
• [x] FCOS

Requirements

• We recommend you to use Anaconda to create a conda environment:

conda create -n detection python=3.6

• Then, activate the environment:

conda activate detection

• Requirements:

pip install -r requirements.txt 

We suggest that PyTorch should be higher than 1.9.0 and Torchvision should be higher than 0.10.3. At least, please make sure your torch is version 1.x.

Main results on COCO-val

YOLOF

RetinaNet

In my RetinaNet:

• For regression head, GIoU Loss is deployed rather than SmoothL1Loss

FCOS

In my FCOS:

• For regression head, GIoU loss is deployed rather than IoU loss
• For real-time FCOS, the PaFPN is deployed for fpn

Train

Single GPU

sh train.sh

You can change the configurations of train.sh, according to your own situation.

Multi GPUs

sh train_ddp.sh

You can change the configurations of train_ddp.sh, according to your own situation.

In the event of a training interruption, you can pass --resume the latest training weight path (None by default) to resume training. For example:

python train.py \
        --cuda \
        -lr 0.01 \
        -lr_bk 0.01 \
        --batch_size 16 \
        --resume weights/coco/yolof18/yolof18_epoch_6_24.8.pth \

Test

python test.py -d coco \
               --cuda \
               -v yolof50 \
               --weight path/to/weight \
               --root path/to/dataset/ \
               --show

Demo

I have provide some images in data/demo/images/, so you can run following command to run a demo:

python demo.py --mode image \
               --path_to_img data/demo/images/ \
               -v yolof50 \
               --cuda \
               --weight path/to/weight \
               --show

If you want run a demo of streaming video detection, you need to set --mode to video, and give the path to video --path_to_vid。

python demo.py --mode video \
               --path_to_img data/demo/videos/your_video \
               -v yolof50 \
               --cuda \
               --weight path/to/weight \
               --show

If you want run video detection with your camera, you need to set --mode to camera。

python demo.py --mode camera \
               -v yolof50 \
               --cuda \
               --weight path/to/weight \
               --show