Core Concepts
Developing a unified model structure for face perception enhances task extensibility and application efficiency.
Abstract
The article introduces Faceptor, a generalist model for face perception that focuses on a unified model structure. It explores shared structural designs and shared parameters to improve task extensibility and application efficiency. The Naive Faceptor consists of one shared backbone and three standardized output heads, while the Faceptor adopts a single-encoder dual-decoder architecture with task-specific queries. The Layer-Attention mechanism is introduced to adaptively select features from optimal layers. Experimental results show exceptional performance in various face analysis tasks, including facial landmark localization, face parsing, age estimation, expression recognition, binary attribute classification, and face recognition.
Stats
Joint training on 13 face perception datasets.
Achieved outstanding performance in various tasks.
Parameters distribution in Naive Faceptor and Faceptor.
Performance comparison between Naive Faceptor and Faceptor.
Quotes
"Existing methods mainly discuss unified representation and training."
"Our contributions can be summarized as follows."
"In multi-task learning, the objective is to achieve optimal performance across all tasks."