Research map / Detection and segmentation / CNN features
Funnel Activation for Visual Recognition
Ningning Ma, Xiangyu Zhang, Jian Sun
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 ReLU with FReLU | ResNet-50 · ImageNet 2012 val | top-1 error (%): 24.0 -> 22.4 | Improved |
| replace ReLU with FReLU | ResNet-101 · ImageNet 2012 val | top-1 error (%): 22.8 -> 22.1 | Improved |
| replace ReLU with FReLU | ShuffleNetV2 · ImageNet 2012 val | top-1 error (%): 39.6 -> 37.1 | Improved |
| use FReLU backbone instead of ReLU | RetinaNet ResNet-50 · COCO minival | mAP: 35.2 -> 36.6 | Improved |
| use FReLU backbone instead of ReLU | PSPNet ResNet-50 · Cityscapes val | mean IU: 77.2 -> 78.9 | Improved |
| use parametric pooling instead of max pooling in the funnel condition (ablation) | FReLU · ImageNet 2012 val | top-1 error (%): 24.4 -> 22.4 | Improved |
| use parametric pooling instead of average pooling in the funnel condition (ablation) | FReLU · ImageNet 2012 val | top-1 error (%): 24.5 -> 22.4 | Improved |
| combine spatial condition and non-linearity with max rather than sum (ablation) | FReLU · ImageNet 2012 val | top-1 error (%): 23.6 -> 22.4 | Improved |
| add max non-linearity on top of the spatial condition (vs depth-wise conv then ReLU) (ablation) | FReLU · ImageNet 2012 val | top-1 error (%): 23.7 -> 22.4 | Improved |
| use 3x3 window instead of 1x1 (PReLU case) (ablation) | FReLU · ImageNet 2012 val | top-1 error (%): 23.7 -> 22.4 | Improved |
| use 3x3 window instead of 5x5 or 7x7 (ablation) | FReLU · ImageNet 2012 val | top-1 error (%): 22.9 -> 22.4 | Improved |
| apply FReLU after both 1x1 and 3x3 convs instead of one (ablation) | FReLU · ImageNet 2012 val | top-1 error (%): 22.9 -> 22.4 | Improved |
Closest papers by meaning
- Higher-Order Integration of Hierarchical Convolutional Activations for Fine-Grained Visual Categorization (ICCV 2017 · similarity 0.84)
- Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR 2021 · similarity 0.83)
- FocSAM: Delving Deeply into Focused Objects in Segmenting Anything (CVPR 2024 · similarity 0.82)
- TrAct: Making First-layer Pre-Activations Trainable (NeurIPS 2024 · similarity 0.82)
- Doomed to Re-Annotate, Forever: The ImageNet Story (NeurIPS 2026 · similarity 0.82)
- Not All Diffusion Model Activations Have Been Evaluated as Discriminative Features (NeurIPS 2024 · similarity 0.82)
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.
