Ant Group and CVPR 2022 Host Dual Competitions on Pet Biometric Recognition and Image Forgery Detection

CVPR, the world‑class IEEE conference on computer vision and pattern recognition, attracts thousands of professionals each year as a premier venue for showcasing cutting‑edge research and exploring real‑world applications.

As a deep‑tech player in artificial intelligence, Ant Group collaborated with the CVPR22 17th IEEE Computer Society Workshop on Biometrics and, together with leading universities and industry institutes, organized two parallel competitions on Alibaba Cloud Tianchi.

The contests drew more than 1,300 registered teams worldwide. Teams from Huazhong University of Science and Technology, ShanghaiTech University and Nanjing University of Science and Technology won the top three places in the "Pet Biometric Recognition" track, while teams from NetEase, ByteDance and Meitu secured the podium in the "Image Forgery Detection" track.

Dual tracks jointly explore computer‑vision innovation

Track 1: Pet Biometric Recognition Competition

With the rapid growth of the pet economy, AI‑enabled pet identity verification has become a hot topic for applications such as pet management, trading, healthcare and insurance. Ant Group and the AI Innovation & Industry Institute of Fudan University launched this competition to encourage solutions for the challenging problem of pet nose‑print identification.

Huazhong University of Science and Technology: Proposed an instance‑feature weighted contrast loss that emphasizes hard negative samples and strengthens hard positive learning, improving fine‑grained discrimination of pet nose‑prints. Data augmentation and multi‑model fusion were also employed to handle lighting and pose variations, achieving an AUC of 0.9087.

ShanghaiTech University: Treated the task as an image‑retrieval problem, using supervised contrast loss and a dual‑global descriptor. They applied progressive image resizing, pooled features at inference to retain detail, and leveraged TTA, AMP and EMA techniques to boost generalisation.

Nanjing University of Science and Technology: Adopted multi‑scale feature fusion with offline random affine, blur and sharpening augmentations, and online Aug‑Mix, Color‑Jitter, etc. Their backbone ResNeSt with Gem Pooling and a BN‑neck head used label‑smoothed cross‑entropy, soft‑margin triplet and circle losses, plus XBM hard‑example mining; post‑processing concatenated multi‑scale features and cosine similarity.

Prof. Qi Yuan, Distinguished Professor at Fudan University, emphasized that the competition aimed to provide realistic data and tasks for top algorithm talent, encouraging innovative solutions rather than simple stacking of existing methods. All winning solutions have been open‑sourced to promote AI adoption in the pet industry.

Track 2: Image Forgery Detection Competition

Advances in image editing have enabled easy creation of forged pictures used for disinformation, pornography and privacy violations. Ant Group, together with Nanyang Technological University (Singapore) and the China Academy of Information and Communications Technology (CAICT), organized this contest to spur research on robust forgery detection. NetEase, ByteDance and Meitu achieved AUC scores of 0.9938, 0.9913 and 0.9879 respectively.

NetEase: Developed DAME (Data Augmentation and Model Ensemble) for generalized face forgery detection, generating ~400 k synthetic fake images via facial reenactment, swapping, attribute editing, cartoonization and artistic filtering. They introduced a hypothesis of consistent real‑image feature distribution, shifting focus from known fake cues to any non‑real features, and built a low‑correlation ensemble of diverse backbones with tailored learning‑rate schedules.

ByteDance: Implemented a multimodal, multi‑class detection framework. In addition to the RGB input, a custom Spatial Rich Model (SRM) filter produced a second modality emphasizing residual noise. The AIM (Augmentation Inside Mask) module performed online fake‑sample generation on masked face regions, simulating diverse forgery artifacts. Multiple models from different training stages were ensembled.

Meitu: Constructed a model pool of seven heterogeneous backbones (Swin‑Transformer, VAN, CDCNpp, Multi‑Attention, ConvNeXt, etc.) each focusing on distinct feature representations (RGB, texture, local patches). PSO‑based search and manual tuning determined ensemble weights. Data diversity was enhanced via self‑blending and local‑cutting augmentations based on face masks, and an extra evaluation set created with FaceMorph, StarGAN, StyleGAN, FaceEdit and FaceSwap was used to assess generalisation.

Chen Tao, Ant Group’s Digital Identity Lead, noted that deep‑fake attacks expose robustness gaps in second‑generation data‑driven AI, while third‑generation trustworthy AI combines data‑driven and knowledge‑driven techniques. Liu Yan, head of Ant Group’s “Sky‑Gap” Lab, highlighted trends such as massive synthetic datasets, ensemble‑based detection, and physiological‑signal‑based features, stressing the importance of interpretability and multi‑modal evidence for real‑world deployment.

Ant Group has long invested in computer‑vision research, applying it across digital life, finance, and security services. This dual‑track competition, conducted in close collaboration with academia and industry, aims to accelerate progress in real‑world visual recognition technologies.

For an in‑depth review of the competition and winning solutions, watch the live summary streamed on June 18 at 19:30 by the Machine‑Heart video channel.