Computation Redistribution: Improving Face Detection with Neural Architecture Search
Computation redistribution is a method used for improving face detection using neural architecture search. Face detection is the ability of a computer program to identify and locate human faces in digital images or videos.
Typically, in computer vision, neural networks are used for this task. These neural networks are made up of different parts, including the backbone, neck, and head of the model. However, when directly utilizing the backbone of a classification network for scale-specific face detection, it can be sub-optimal.
Therefore, network structure search is used to reallocate the computation on the backbone, neck, and head, under a wide range of flop regimes. This optimization allows for better face detection results.
How Computation Redistribution Works
The search method called computation redistribution is applied to RetinaNet, a neural network with ResNet as the backbone, Path Aggregation Feature Pyramid Network (PAFPN) as the neck, and stacked 3 × 3 convolutional layers for the head.
In the first step, the authors explore the reallocation of the computation within the backbone parts (i.e., stem, C2, C3, C4, and C5), while fixing the neck and head components. Based on the optimized computation distribution on the backbone they find, they further explore the reallocation of the computation across the backbone, neck, and head.
The Results of Computation Redistribution
Using computation redistribution, the search methodology identified the optimal computation distribution for RetinaNet, improving its face detection accuracy across a wide range of computational resources. The search space was large, but computation redistribution allowed for a more efficient exploration of that space.
This improvement in accuracy means that face detection software can better identify human faces in images and videos, leading to more accurate and efficient algorithms for various applications, including security systems and video analytics.
Future Applications of Computation Redistribution
Computation redistribution could be used for other computer vision tasks beyond face detection. Other applications in the field of computer vision may benefit from this optimization. It could improve the performance of neural networks and enhance the accuracy of various computer vision tasks.
In the future, computation redistribution may become a standard technique for improving neural networks for computer vision applications.
