https://github.com/grapeot/AnimeHeadDetector

An object detector for character heads in animes, based on Yolo V3.
https://github.com/grapeot/AnimeHeadDetector

Last synced: about 1 month ago
JSON representation

An object detector for character heads in animes, based on Yolo V3.

• Host: GitHub
• URL: https://github.com/grapeot/AnimeHeadDetector
• Owner: grapeot
• License: mit
• Created: 2019-06-26T21:01:16.000Z (about 5 years ago)
• Default Branch: master
• Last Pushed: 2019-07-03T17:19:02.000Z (almost 5 years ago)
• Last Synced: 2024-04-14T20:00:48.025Z (2 months ago)
• Language: Python
• Size: 622 KB
• Stars: 94
• Watchers: 2
• Forks: 11
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Lists

• awesome-face-detection-and-recognition - grapeot/AnimeHeadDetector
• awesome-yolo-object-detection - grapeot/AnimeHeadDetector

README

        
# AnimeHeadDetector

An object detector for character heads in animes, based on Yolo V3.

## Motivation

Yet another anime face/head detector, in addition to [Faster RCNN](https://github.com/qhgz2013/anime-face-detector) and [Viola Detector](https://github.com/nagadomi/lbpcascade_animeface).

The major difference is this one is more on character head instead of face only, so we also have outputs when the face is not visible.

The algorithm is based on [Yolo](https://pjreddie.com/darknet/yolo/) V3 and [Pytorch implementation](https://github.com/ayooshkathuria/pytorch-yolo-v3) so it's good for high-speed processing and the usage is straight-forward.

## Results

Sample results:

![](https://raw.githubusercontent.com/grapeot/AnimeHeadDetector/master/assets/57032.jpg)

We can detect correctly when a hat is present, but uncommon expression + decoration will fail the detection.

![](https://raw.githubusercontent.com/grapeot/AnimeHeadDetector/master/assets/112032.jpg)

![](https://raw.githubusercontent.com/grapeot/AnimeHeadDetector/master/assets/204032.jpg)

![](https://raw.githubusercontent.com/grapeot/AnimeHeadDetector/master/assets/164032.jpg)

We can handle the case where the face is barely visible.

![](https://raw.githubusercontent.com/grapeot/AnimeHeadDetector/master/assets/119032.jpg)

Although all our training data was labeled on a color anime set, the algorithm generalizes well on line drawings.

## Usage

* First clone the repo using `git clone --recursive https://github.com/grapeot/AnimeHeadDetector`. Note the `--recursive` switch. It's necessary because we used `git submodule` to manage the dependency of PyTorch implementation.

* Use `pip install -r requirements.txt` to install the dependencies.

* And then download the pre-trained model using the `./downloadWeights.sh` script. (I just found my website cannot be accessed from within China. For China users, you can also download using this [Baidu Pan link](https://pan.baidu.com/s/1Ke0DZ7Ptai5RCc3wL1VDMA), with access code `m88e`.)

* `main.py` provides all necessary functionality. `python main.py` should run out of box. It takes the images in `imgs`, does detection, and visualizes the results in `det`. Check `python main.py --help` for detailed usage.

## More comments

* You may want to tweak the `CONFIDENCE_THRESHOLD` in `AnimeHeadDetector.py` for the desired detection precision/recall.

* The training data is merely 200 images from two animes. The algorithm works surprisingly well.