Research map / Image generation / Image restoration
Joint Image Filtering with Deep Convolutional Networks
Yijun Li, Jia‐Bin Huang, Narendra Ahuja, Ming–Hsuan Yang
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 |
|---|---|---|---|
| add skip connection (residual learning) to the joint filter network (ablation) | DJF · NYU v2 depth upsampling 8x | RMSE (cm): 6.20 -> 5.86 | Improved |
| add skip connection at large filter number n1=256 (ablation) | joint filter network · NYU v2 depth upsampling 8x | RMSE (cm): 6.40 -> 5.82 | Improved |
| increase filter number from n1=64 to n1=256 with skip connection (ablation) | joint filter network with skip · NYU v2 depth upsampling 8x | RMSE (cm): 5.97 -> 5.82 | Improved |
| increase filter number without skip connection (ablation) | joint filter network without skip · NYU v2 depth upsampling 8x | RMSE (cm): 6.40, 6.44, 6.32, 6.35 | No clear effect |
| use multi-channel output features n3=16 instead of n3=1 (ablation) | CNN_T and CNN_G with skip · NYU v2 depth upsampling 8x | RMSE (cm): 5.86 -> 6.11 | Hurt |
| increase filter sizes from f=5/1/3 to f=9/1/5 (ablation) | joint filter network with skip · NYU v2 depth upsampling 8x | RMSE (cm): 6.24 -> 5.86 | Improved |
| enlarge filter size further to f=11/3/7 (ablation) | joint filter network with skip · NYU v2 depth upsampling 8x | RMSE (cm): 5.86 -> 5.93 | No clear effect |
| increase subnetwork depth from d=2 to d=5 (ablation) | proposed model · NYU v2 depth upsampling 8x | RMSE (cm): 5.99 -> 5.73 | Improved |
| deepen a single generic residual network CNN_F_R from d=3 to d=8 (ablation) | CNN_F_R · NYU v2 depth upsampling 8x | RMSE (cm): 6.31 -> 6.16 | Improved |
| add separate target and guidance sub-networks instead of stacking inputs (ablation) | proposed model · NYU v2 depth upsampling 8x | RMSE (cm): 6.13 -> 5.86 | Improved |
| make CNN_F shallower with 4/4-2 depth split (ablation) | proposed model · NYU v2 depth upsampling 8x | RMSE (cm): 5.86 -> 6.03 | Hurt |
| train on RGB/depth instead of RGB/flow (ablation) | proposed model · NYU v2 depth upsampling 8x | RMSE (cm): 7.32 -> 5.86 | Improved |
Closest papers by meaning
- Learning Spatially Variant Linear Representation Models for Joint Filtering (TPAMI 2021 · similarity 0.93)
- Spatially Variant Linear Representation Models for Joint Filtering (CVPR 2019 · similarity 0.93)
- Fast End-to-End Trainable Guided Filter (CVPR 2018 · similarity 0.89)
- Deformable Kernel Networks for Joint Image Filtering (IJCV 2020 · similarity 0.88)
- Deep Joint Image Filtering (ECCV 2016 · similarity 0.87)
- Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning (NeurIPS 2025 · 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.
