vcan / AniSeg

A faster-rcnn model for anime character segmentation.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

AniSeg

Tensorflow models for Anime character object detection.

Samples

Face detection

Face detection sample 1

Face detection sample 2

Figure segmentation

Figure detection sample 1

Figure detection sample 2

Figure detection sample 3

Usage

We provide two pretrained models: One for face detection and one for figure segmentation. Download them at Google Drive!

Here is a full example script to classify faces in the sample images:

# Assume you have saved the pretrained models under "model/" directory.

python infer_from_image.py \
--input_images=samples/inputs/*
--output_path=samples/face_detection_outputs
--inference_graph=model/face_detection/frozen_inference_graph.pb
--visualize_inference=True

In general, use the script with the following format

python infer_from_image.py \
--input_images=/PATH/TO/IMAGES/* \
--output_path=/PATH/TO/OUTPUT/FOLDER/  \
--inference_graph=/PATH/TO/frozen_inference_graph.pb  \
--visualize_inference=True

For figure segmentation model, please point --inference_graph to the segmentation model and add the additional flag

--detect_masks=True

About the models

The Face Detection is trained on the outputs of animeface-2009 detector, which is under the MIT license. The main motivation for a tensorflow version is for a much faster GPU-based detection (5-10x speedup).

The Figure Segmentation model is trained on a manually annotated dataset, which you can download here. Note some images are NSFW. We release this dataset purely for research. Please use at your own risk.

To train the model, we overlayed segmented anime figures on top of pure background images to create an artificial dataset. We found this gives a decent performance. You can find pure background images in the Danbooru 2018 dataset. Please contact us if you'd like to use our pre-generated tfrecords.

Both model is trained using the Tensorflow Object Detection repo. Our repo is a simplified fork of that.

Error cases

  • Duplicate objects

Detection error 1

  • False negatives (missing detection)

Detection error 2

  • False positives

Detection error 3

Anime related repos and datasets

Shameless self promotion of my TwinGAN model to turn people into anime characteres and cats!

Sketch coloring using PaintsTransfer and PaintsChainer.

Create anime portraits at Crypko and MakeGirlsMoe

The all-encompassing anime dataset Danbooru2017 by gwern.

My hand-curated sketch-colored image dataset.

Common questions

What's the use case?

Getting clean data is hard. The object detectors provided here can make that process easier. For example it can be used to crop faces from the Danbooru2017 dataset. Here we show some sample results of combining the two mdoels:

use case 1 use case 2 use case 3 use case 4 use case 5

Results are too noisy

There is always a tradeoff between precision and recall. Please adjust the --min_score_thresh flag accordingly.

Higher threshold means more accurate results but also more false negatives (higher chance to miss objects). Lower value gives more faces but also contains non-face objects.

Inference is too slow

Inference using CPU is slow. Please make sure you have a GPU/TPU.

Citation

If you decide to use our model and/or data, we ask you to kindly cite this repo. Have fun!

About

A faster-rcnn model for anime character segmentation.

License:Apache License 2.0


Languages

Language:Python 99.6%Language:Shell 0.4%