CCIP
CCIP(Contrastive Anime Character Image Pre-Training) is a model to calculuate the visual similarity between anime characters in two images. (limited to images containing only a single anime character). More similar the characters between two images are, higher score it should have.
Usage
Using CCIP with imgutils
Calculuate character similarity between images:
from imgutils.metrics import ccip_batch_differences
ccip_batch_differences(['ccip/1.jpg', 'ccip/2.jpg', 'ccip/6.jpg', 'ccip/7.jpg'])
array([[6.5350548e-08, 1.6583106e-01, 4.2947042e-01, 4.0375218e-01],
[1.6583106e-01, 9.8025822e-08, 4.3715334e-01, 4.0748104e-01],
[4.2947042e-01, 4.3715334e-01, 3.2675274e-08, 3.9229470e-01],
[4.0375218e-01, 4.0748104e-01, 3.9229470e-01, 6.5350548e-08]],
dtype=float32)
Performence
Model | F1 Score | Precision | Recall | Threshold | Cluster_2 | Cluster_Free |
---|---|---|---|---|---|---|
ccip-caformer_b36-24 | 0.940925 | 0.938254 | 0.943612 | 0.213231 | 0.89508 | 0.957017 |
ccip-caformer-24-randaug-pruned | 0.917211 | 0.933481 | 0.901499 | 0.178475 | 0.890366 | 0.922375 |
ccip-v2-caformer_s36-10 | 0.906422 | 0.932779 | 0.881513 | 0.207757 | 0.874592 | 0.89241 |
ccip-caformer-6-randaug-pruned_fp32 | 0.878403 | 0.893648 | 0.863669 | 0.195122 | 0.810176 | 0.897904 |
ccip-caformer-5_fp32 | 0.864363 | 0.90155 | 0.830121 | 0.183973 | 0.792051 | 0.862289 |
ccip-caformer-4_fp32 | 0.844967 | 0.870553 | 0.820842 | 0.18367 | 0.795565 | 0.868133 |
ccip-caformer_query-12 | 0.823928 | 0.871122 | 0.781585 | 0.141308 | 0.787237 | 0.809426 |
ccip-caformer-23_randaug_fp32 | 0.81625 | 0.854134 | 0.781585 | 0.136797 | 0.745697 | 0.8068 |
ccip-caformer-2-randaug-pruned_fp32 | 0.78561 | 0.800148 | 0.771592 | 0.171053 | 0.686617 | 0.728195 |
ccip-caformer-2_fp32 | 0.755125 | 0.790172 | 0.723055 | 0.141275 | 0.64977 | 0.718516 |
- The calculation of
F1 Score
,Precision
, andRecall
considers "the characters in both images are the same" as a positive case.Threshold
is determined by finding the maximum value on the F1 Score curve. Cluster_2
represents the approximate optimal clustering solution obtained by tuning the eps value in DBSCAN clustering algorithm with min_samples set to2
, and evaluating the similarity between the obtained clusters and the true distribution using therandom_adjust_score
.Cluster_Free
represents the approximate optimal solution obtained by tuning themax_eps
andmin_samples
values in the OPTICS clustering algorithm, and evaluating the similarity between the obtained clusters and the true distribution using therandom_adjust_score
.
Citation
@misc{CCIP,
title={Contrastive Anime Character Image Pre-Training},
author={Ziyi Dong and narugo1992},
year={2024},
howpublished={\url{https://huggingface.co./deepghs/ccip}}
}
Inference API (serverless) does not yet support dghs-imgutils models for this pipeline type.