Related Experiment Videos
PicassoNet: Searching Adaptive Architecture for Efficient Facial Landmark Localization
IEEE Transactions on Neural Networks and Learning Systems
|April 28, 2022
Summary
PicassoNet, a lightweight facial landmark detector, achieves fast inference speeds and comparable accuracy by adaptively allocating computation to facial parts. This novel approach enhances efficiency for real-world facial applications.
Area of Science:
- Computer Vision
- Machine Learning
- Deep Learning
Background:
- Facial landmark localization is crucial for many applications.
- Existing methods often sacrifice speed for accuracy.
- Efficient facial landmark detection remains a challenge.
Purpose of the Study:
- To develop a lightweight and efficient facial landmark detector.
- To improve inference speed without compromising accuracy.
- To introduce adaptive computation for individual facial parts.
Main Methods:
- Proposed PicassoNet, a cascaded detector with adaptive computation using group convolution.
- Integrated refinement submodules into a single network.
- Utilized a novel grouping search algorithm, formulated as a network architecture search (NAS) problem, for adaptive computation allocation.
- Introduced a boundary-aware loss function for improved keypoint localization.
Main Results:
- PicassoNet achieved over 6x faster inference speed compared to state-of-the-art methods.
- The method maintained comparable accuracy on benchmark datasets (AFLW, 300W, WFLW).
- Adaptive computation allocation led to an efficient network structure.
Conclusions:
- PicassoNet offers a significant improvement in efficiency for facial landmark detection.
- The adaptive computation and boundary-aware loss contribute to both speed and accuracy.
- This method is well-suited for real-time facial analysis applications.