ACFD: Asymmetric Cartoon Face Detector

Bin Zhang, Jian Li, Yabiao Wang, Zhipeng Cui, Yili Xia, Chengjie Wang, Jilin Li, Feiyue HuangarXiv 2020 · arXiv

Introduction

Cartoon face detection presents unique challenges compared to human face detection due to high intra-face diversity, extreme aspect ratios, and background confusion. ACFD (Asymmetric Cartoon Face Detector) is a one-stage framework designed to address these hurdles. It achieved 1st place in the 2020 iCartoon Face Challenge by leveraging asymmetric architectural components and dynamic training strategies.

Architecture and Methodology

The ACFD framework consists of three main components: a feature extraction backbone, a multi-scale feature fusion neck, and a prediction head.

The overall ACFD architecture featuring VoVNetV3-51, ABi-FPN, and Anchor Heads.

VoVNetV3 and Asymmetric Convolution

The backbone, VoVNetV3, evolves from the One-Shot Aggregation (OSA) architecture. The core innovation is the Asymmetric One-Shot Aggregation (AOSA) module, which incorporates Asymmetric Convolution Blocks (ACB). An ACB replaces standard $3 \times 3$ convolutions with a parallel combination of $3 \times 3$, $1 \times 3$, and $3 \times 1$ kernels. This design allows the network to capture features with diverse, non-square receptive fields, making it robust to the exaggerated proportions common in cartoon art.

Detailed view of the AOSA module and the Asymmetric Convolution Block (ACB) used within the backbone.

ABi-FPN

The Asymmetric Bi-directional Feature Pyramid Network (ABi-FPN) serves as the neck module. Similar to the backbone, it utilizes ACBs to fuse multi-scale features effectively, particularly for faces with extreme poses and aspect ratios.

Training Innovations

Beyond architectural changes, ACFD introduces two critical training strategies:

  1. Dynamic Anchor Match (DAM): Unlike static IoU-based matching, DAM evaluates the regression potential of anchors. If an anchor's predicted box shows high overlap with the ground truth during training, it is treated as a positive sample, even if the initial anchor overlap was low. This was the most impactful algorithmic change, contributing a +1.25% mAP gain.
  2. Margin Binary Classification Loss (MBC): By incorporating a hard margin into the classification loss (inspired by ArcFace), the model learns more discriminative features, reducing false positives in cluttered backgrounds.

Results and Performance

ACFD won the 2020 iCartoon Face Challenge with a 92.91% mAP. Ablation studies highlight the effectiveness of its components:

  • VoVNetV3: +0.57% mAP over VoVNetV2.
  • DAM: +1.25% mAP (most significant gain).
  • ABi-FPN: +0.18% mAP over standard BiFPN.

The model is highly efficient, with a size under 200MB and an inference time of less than 50ms per image on a standard GPU.

Conclusion

ACFD demonstrates that explicitly modeling asymmetry in receptive fields and dynamically matching anchors can significantly improve detection performance in stylized domains like cartoon face recognition. Its success in competition validates these design choices for real-world deployment.

Summary updated

Abstract

Cartoon face detection is a more challenging task than human face detection due to many difficult scenarios is involved. Aiming at the characteristics of cartoon faces, such as huge differences within the intra-faces, in this paper, we propose an asymmetric cartoon face detector, named ACFD. Specifically, it consists of the following modules: a novel backbone VoVNetV3 comprised of several asymmetric one-shot aggregation modules (AOSA), asymmetric bi-directional feature pyramid network (ABi-FPN), dynamic anchor match strategy (DAM) and the corresponding margin binary classification loss (MBC). In particular, to generate features with diverse receptive fields, multi-scale pyramid features are extracted by VoVNetV3, and then fused and enhanced simultaneously by ABi-FPN for handling the faces in some extreme poses and have disparate aspect ratios. Besides, DAM is used to match enough high-quality anchors for each face, and MBC is for the strong power of discrimination. With the effectiveness of these modules, our ACFD achieves the 1st place on the detection track of 2020 iCartoon Face Challenge under the constraints of model size 200MB, inference time 50ms per image, and without any pretrained models.
BibTeX
@misc{zhang2020acfd,
  author = {Bin Zhang and Jian Li and Yabiao Wang and Zhipeng Cui and Yili Xia and Chengjie Wang and Jilin Li and Feiyue Huang},
  title = {{ACFD: Asymmetric Cartoon Face Detector}},
  year = {2020},
  eprint = {2007.00899},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2007.00899},
}
Figure