A Pose Estimation Method for Multiple Identity Cards based on Corner Heatmaps and Part Affinity Fields

Automatic information extraction from identity cards is crucial in many applications such as eKYC, customer registration, and profile digitalization. Identity card detection and normalization are crucial steps before further information extraction. In practice, many identity cards of different types with both front and back sides may appear in an image or on the same page of a profile. This paper proposes a method to estimate the pose of multiple identity cards and classify the card into four categories: front/back sides of new/old identity cards. Particularly, we propose an EfficientNet-based model to detect the corners in an anchor-free manner and estimate the edges of identity cards using part affinity fields (PAF). According to the estimated PAF, the detected corners are then grouped in clusters, each of which corresponds to a single identity card. Experimental results show that our method yields promising results on our private dataset. We achieve a realtime speed of 24 fps on a machine with NVIDIA GPU GTX 1080 Ti.