Research map / Architectures and efficiency / CNN design
Refining Activation Downsampling With SoftPool
Alexandros Stergiou, Ronald Poppe, Grigorios Kalliatakis
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 |
|---|---|---|---|
| replace max pooling with SoftPool, trained from scratch | ResNet18 · ImageNet1K val (from scratch) | top-1 accuracy: 69.76 -> 71.27 | Improved |
| replace max pooling with SoftPool, trained from scratch | ResNeXt101 32x8d · ImageNet1K val (from scratch) | top-1 accuracy: 79.31 -> 80.12 | Improved |
| swap in SoftPool on a pre-trained network without retraining | ResNet18 · ImageNet1K val (pre-trained, pooling swapped) | top-1 accuracy: 69.76 -> 70.56 | Improved |
| train from scratch with SoftPool instead of only swapping it into a pre-trained network | ResNet18 · ImageNet1K val | top-1 accuracy: 70.56 -> 71.27 | Improved |
| replace pooling with SoftPool rather than LIP | ResNet18 · ImageNet1K val (from scratch, pooling comparison) | top-1 accuracy: 70.83 -> 71.27 | Improved |
| replace pooling with SoftPool rather than stochastic pooling | ResNet18 · ImageNet1K val (from scratch, pooling comparison) | top-1 accuracy: 70.13 -> 71.27 | Improved |
| progressively replace all seven max-pool stages with SoftPool (ablation) | InceptionV3 · ImageNet1K val (InceptionV3 progressive replacement) | top-1 accuracy: 77.45 -> 79.04 | Improved |
| replace only the first pooling layer with SoftPool (ablation) | InceptionV3 · ImageNet1K val (InceptionV3 progressive replacement) | top-1 accuracy: 77.45 -> 77.93 | Improved |
| replace pooling with SoftPool | r3d-50 · HACS | top-1 accuracy: 78.36 -> 79.82 | Improved |
| replace pooling with SoftPool | SRTG r(2+1)d-50 · HACS | top-1 accuracy: 83.77 -> 84.78 | Improved |
| replace pooling with SoftPool | SRTG r3d-101 · Kinetics-700 (fine-tuned) | top-1 accuracy: 56.46 -> 57.76 | Improved |
| replace pooling with SoftPool | SRTG r3d-101 · UCF-101 (fine-tuned) | top-1 accuracy: 97.32 -> 98.06 | Improved |
Closest papers by meaning
- LiftPool: Bidirectional ConvNet Pooling (ICLR 2021 · similarity 0.88)
- S3Pool: Pooling With Stochastic Spatial Sampling (CVPR 2017 · similarity 0.88)
- AdaScan: Adaptive Scan Pooling in Deep Convolutional Neural Networks for Human Action Recognition in Videos (CVPR 2017 · similarity 0.87)
- Gaussian-Based Pooling for Convolutional Neural Networks (NeurIPS 2019 · similarity 0.87)
- Detail-Preserving Pooling in Deep Networks (CVPR 2018 · similarity 0.87)
- Beyond Skip Connection: Pooling and Unpooling Design for Elimination Singularities (AAAI 2025 · similarity 0.84)
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.
