Dynamic Filtering with Large Sampling Field for ConvNets

By Jialin Wu, Daiyang Li, Yu Yang, Chandrajit L. Bajaj, Xiangyang Ji. 2018

43 citations recorded by the source corpus.

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 neighbour 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 [5], namely avoiding feature map blurring by positionwise 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 [8] and sharper responses in flow estimation on FlyingChairs dataset [6] compared to strong baselines.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Optimization & Training

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Optimization & Training

Citations in the corpus

K-Net: Towards Unified Image Segmentation (2021, methodology): “Some semantic segmentation methods apply dynamic kernels to improve the model representation with enlarged receptive fields [56] or multi-scales contexts [20].”
Dynamic Neural Networks: A Survey (2021, background): “2), pixel-level dynamic weights are achieved by test-time adjustment [108], [109], prediction [151], [152], [153], [154] or dynamic features [21], [129], [130], [135].”
Dynamic Multi-Scale Filters for Semantic Segmentation (2019, methodology): “DFN [23] exploits dynamic filters to capture different motion patterns within the inputs for video and stereo prediction, and [41] also uses dynamic filters aiming at constructing large receptive fields and receiving local gradients, to produce sharper and more semantic feature maps.”
Dynamic Fusion Module Evolves Drivable Area and Road Anomaly Detection: A Benchmark and Algorithms (2021, methodology): “[34] developed an extension of the DFN by dynamically generating weights to enlarge receptive fields for flow estimation.”
Learning Depth-Guided Convolutions for Monocular 3D Object Detection (2019, methodology): “Other techniques like deformable convolution [8] and variants of [20] such as [14, 46, 52, 11], fail to capture object scale and local structure as well.”
Faithful Multimodal Explanation for Visual Question Answering (2018, methodology): “As recent research suggested that gradientbased methods more faithfully present the models’ decision making process (Zhang et al., 2018; Wu et al., 2018b; Wu and Mooney, 2019; Jain and Wallace, 2019), we define a faithfulness score Sf as the cosine similarity between the Grad-CAM (Selvaraju et al.”
Learning to Predict Layout-to-image Conditional Convolutions for Semantic Image Synthesis (2019, methodology): “This idea has been applied to different applications such as neural style transfer [27], superresolution [15, 12], image segmentation [8, 30], motion prediction [32] and tracking [19].”
Hyper-LifelongGAN: Scalable Lifelong Learning for Image Conditioned Generation (2021, methodology): “This idea has been applied to applications in different research fields such as few shot learning [27, 33], image segmentation [2, 36] and generative models [28, 18, 16], which is the focus of our paper.”
Learning Guided Convolutional Network for Depth Completion (2019, background): “Such an average is more likely leading to gradient closing to zero, even though the learned kernel is far from optimal for every position, which could generate sub-optimal results as pointed by [48].”