Jialin Wu; Dai Li; Yu Yang; Bajaj, C.; Xiangyang Ji. European Conference on Computer Vision (ECCV),2018.时间: 2018-09-08 点击: 398 次 返回列表
We present Dynamic Sampling Convolutional Neural Networks (DSCNN), where the position-specific kernels learn from not only the current position but also multiple sampled neighbour regions.
We propose a dynamic filtering strategy with large samplingfield for ConvNets (LS-DFN), where the position-specific kernels learnfrom not only the identical position but also multiple sampled neighbourregions. During sampling, residual learning is introduced to ease trainingand an attention mechanism is applied to fuse features from differentsamples. Such multiple samples enlarge the kernels receptive fields significantlywithout requiring more parameters. While LS-DFN inherits theadvantages of DFN , namely avoiding feature map blurring by positionwisekernels while keeping translation invariance, it also efficientlyalleviates the overfitting issue caused by much more parameters thannormal CNNs. Our model is efficient and can be trained end-to-end viastandard back-propagation. We demonstrate the merits of our LS-DFNon both sparse and dense prediction tasks involving object detection,semantic segmentation and flow estimation. Our results show LS-DFNenjoys stronger recognition abilities in object detection and semantic segmentationtasks on VOC benchmark  and sharper responses in flowestimation on FlyingChairs dataset  compared to strong baselines.