Research map / Architectures and efficiency / CNN design
Incorporating Side Information by Adaptive Convolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| adaptive convolution with side information (ACNN-v3, all conv layers adaptive) instead of plain CNN | crowd counting on UCSD · UCSD max split | MAE: 1.26 -> 0.96 | Helped |
| adapting only the last conv layer (ACNN-v2) vs only first layer (ACNN-v1) (ablation) | crowd counting on UCSD · UCSD max split | MAE: v2 1.14 vs v1 1.23 | Mixed |
| side information as extra input channel (CNN-X) vs plain CNN | crowd counting on UCSD · UCSD max split | MAE: 1.26 -> 1.20 | Helped |
| adding camera angle, height and perspective as adaptive-conv side information | ACNN on CityUHK-X · CityUHK-X | MAE: 8.72 -> 8.00 | Helped |
| adding more side information (perspective value to angle and height) (ablation) | ACNN on CityUHK-X · CityUHK-X | MAE: 8.35 -> 8.00 | Helped |
| adaptive convolution with noise level as side information (2-conv) | MNIST salt-and-pepper digit recognition · MNIST salt-and-pepper, 2-conv | error rate (%): 8.66 -> 7.55 | Helped |
| adaptive convolution with noise level as side information (4-conv) | MNIST salt-and-pepper digit recognition · MNIST salt-and-pepper, 4-conv | error rate (%): 3.58 -> 2.92 | Helped |
| side info as extra input channel (CNN-X) for digit recognition (ablation) | MNIST salt-and-pepper digit recognition · MNIST salt-and-pepper, 4-conv | error rate (%): 3.58 -> 3.57 | No clear effect |
| using estimated noise level instead of ground-truth for ACNN (ablation) | MNIST salt-and-pepper digit recognition · MNIST salt-and-pepper, 4-conv | error rate (%): 2.92 -> 2.97 | Hurt |
| adaptive convolution with kernel radius as side information (full training radii) | deblurring on Flickr8k · Flickr8k deblurring, trained on r in {3,5,7,9,11} | PSNR gain over blurred input (dB): +0.56 -> +1.03 | Helped |
| adaptive convolution with kernel radius as side information (reduced radii {3,7,11}) | deblurring on Flickr8k · Flickr8k deblurring, trained on r in {3,7,11} | PSNR gain over blurred input (dB): +0.41 -> +0.84 | Helped |
| side info as extra input channel (CNN-X) vs ACNN for deblurring on seen radii (ablation) | deblurring on Flickr8k · Flickr8k deblurring, trained on r in {3,5,7,9,11} | PSNR gain over blurred input (dB): CNN-X +1.41 vs ACNN +1.03 on seen radii; CNN-X lower on unseen radii | Mixed |
Closest papers by meaning
- Context-Aware Crowd Counting (CVPR 2019 · similarity 0.90)
- Single-Image Crowd Counting via Multi-Column Convolutional Neural Network (CVPR 2016 · similarity 0.89)
- Revisiting Perspective Information for Efficient Crowd Counting (CVPR 2019 · similarity 0.89)
- Switching Convolutional Neural Network for Crowd Counting (CVPR 2017 · similarity 0.88)
- Wide-Area Crowd Counting via Ground-Plane Density Maps and Multi-View Fusion CNNs (CVPR 2019 · similarity 0.88)
- Top-Down Feedback for Crowd Counting Convolutional Neural Network (AAAI 2018 · similarity 0.88)
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.