Dynamic Filtering with Large Sampling Field for ConvNets
This work addresses a bottleneck in ConvNets for computer vision tasks like object detection and segmentation, offering an incremental improvement over existing dynamic filtering methods.
The paper tackled the problem of limited receptive fields in ConvNets by proposing a dynamic filtering strategy with large sampling field (LS-DFN), which uses position-specific kernels from multiple neighbor regions to enlarge receptive fields without extra parameters, resulting in stronger recognition abilities in object detection and semantic segmentation on VOC benchmark and sharper responses in flow estimation on FlyingChairs dataset.
We propose a dynamic filtering strategy with large sampling field for ConvNets (LS-DFN), where the position-specific kernels learn from not only the identical position but also multiple sampled neighbor regions. During sampling, residual learning is introduced to ease training and an attention mechanism is applied to fuse features from different samples. Such multiple samples enlarge the kernels' receptive fields significantly without requiring more parameters. While LS-DFN inherits the advantages of DFN, namely avoiding feature map blurring by position-wise kernels while keeping translation invariance, it also efficiently alleviates the overfitting issue caused by much more parameters than normal CNNs. Our model is efficient and can be trained end-to-end via standard back-propagation. We demonstrate the merits of our LS-DFN on both sparse and dense prediction tasks involving object detection, semantic segmentation, and flow estimation. Our results show LS-DFN enjoys stronger recognition abilities in object detection and semantic segmentation tasks on VOC benchmark and sharper responses in flow estimation on FlyingChairs dataset compared to strong baselines.