Research map / Architectures and efficiency / CNN design
Spatially Adaptive Inference with Stochastic Feature Sampling and Interpolation
Zhenda Xie, Zheng Zhang, Xizhou Zhu, Gao Huang, Stephen Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use learnable RBF interpolation kernel instead of average pooling (ablation) | Ours (ResNet-101) · COCO2017 val detection | GFLOPs: 110.0 -> 96.6 | Improved |
| use learnable RBF interpolation kernel instead of plain convolution (ablation) | Ours (ResNet-101) · COCO2017 val detection | GFLOPs: 109.8 -> 96.6 | Improved |
| add interpolation instead of filling zeros (ablation) | Ours (ResNet-101) · COCO2017 val detection | GFLOPs: 164.9 -> 96.6 | Improved |
| add interpolation instead of reusing features (ablation) | Ours (ResNet-101) · COCO2017 val detection | GFLOPs: 226.1 -> 96.6 | Improved |
| add grid prior (stride 11) (ablation) | Ours (ResNet-101) · COCO2017 val detection | mAP: 41.8 -> 42.0 | No clear effect |
| add grid prior at loss weight 0.2 (ablation) | Ours (ResNet-101) · COCO2017 val detection | mAP: 32.63 -> 40.69 | Improved |
| add grid prior at loss weight 0.02 (ablation) | Ours (ResNet-101) · COCO2017 val detection | mAP: 42.61 -> 43.28 | Improved |
| add grid prior to reduce test-time variance (ablation) | Ours (ResNet-101) · COCO2017 val detection | mAP std over 5 tests: 0.09 -> 0.02 | Improved |
| use stochastic Gumbel-Softmax sampling instead of deterministic Gumbel-Softmax (ablation) | Ours (ResNet-101) · COCO2017 val detection | GFLOPs at 42.0 mAP: 121.5 -> 96.6 | Improved |
| use stochastic sampling instead of ReLU gating (ablation) | Ours (ResNet-101) · COCO2017 val detection | GFLOPs at ~42 mAP: 181.9 -> 96.6 | Improved |
| use sampling-interpolation instead of baseline without sampling | ResNet-101 detector · COCO2017 val detection | mAP / GFLOPs: 43.4 mAP 289.5 GFLOPs -> 43.3 mAP 160.4 GFLOPs | Improved |
| use sampling-interpolation on Cityscapes | Segmentation baseline · Cityscapes val segmentation | mean IoU / GFLOPs: 80.8 mean IoU 920.6 GFLOPs -> 80.6 mean IoU 373.2 GFLOPs | Improved |
Closest papers by meaning
- Bottom-Up and Top-Down Reasoning With Hierarchical Rectified Gaussians (CVPR 2016 · similarity 0.86)
- Dynamic Spatial Sparsification for Efficient Vision Transformers and Convolutional Neural Networks (TPAMI 2023 · similarity 0.86)
- Inference, Learning and Attention Mechanisms that Exploit and Preserve Sparsity in CNNs (IJCV 2020 · similarity 0.85)
- Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR 2021 · similarity 0.85)
- Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks With Octave Convolution (ICCV 2019 · similarity 0.85)
- S3Pool: Pooling With Stochastic Spatial Sampling (CVPR 2017 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
