RF-Next: Efficient Receptive Field Search for Convolutional Neural Networks
This addresses the need for automated receptive field optimization in CNNs for researchers and practitioners in computer vision and speech processing, offering a novel search method but with incremental improvements over existing hand-designed approaches.
The paper tackles the problem of manually designing receptive fields in convolutional neural networks by proposing RF-Next, an efficient global-to-local search scheme to automatically find better receptive field combinations. The result shows that RF-Next boosts performance on tasks like temporal action segmentation, object detection, instance segmentation, and speech synthesis, with the source code made publicly available.
Temporal/spatial receptive fields of models play an important role in sequential/spatial tasks. Large receptive fields facilitate long-term relations, while small receptive fields help to capture the local details. Existing methods construct models with hand-designed receptive fields in layers. Can we effectively search for receptive field combinations to replace hand-designed patterns? To answer this question, we propose to find better receptive field combinations through a global-to-local search scheme. Our search scheme exploits both global search to find the coarse combinations and local search to get the refined receptive field combinations further. The global search finds possible coarse combinations other than human-designed patterns. On top of the global search, we propose an expectation-guided iterative local search scheme to refine combinations effectively. Our RF-Next models, plugging receptive field search to various models, boost the performance on many tasks, e.g., temporal action segmentation, object detection, instance segmentation, and speech synthesis. The source code is publicly available on http://mmcheng.net/rfnext.