Research map / Detection and segmentation / CNN features
DenseNets Reloaded: Paradigm Shift Beyond ResNets and ViTs
DongHyun Kim, Byeongho Heo, Dongyoon Han
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 concatenation instead of additive shortcuts in random networks (ablation) | RandNet (Tiny-ImageNet) · Tiny-ImageNet random networks | top-1 accuracy: 55.9 -> 56.8 (averaged over 15k networks) | Improved |
| add modernized feature-mixer blocks (LN, depthwise conv, fewer norms/acts, kernel 7) (ablation) | wide shallow DenseNet · ImageNet-1K val | top-1 accuracy: 79.5 -> 80.4 | Improved |
| widen (GR 32 to 120) and make DenseNet-201 shallower (ablation) | DenseNet-201 · ImageNet-1K val | top-1 accuracy / training speed: 79.7 -> 79.5 acc; train memory 3.9 -> 3.2 GB | Mixed |
| decouple expansion ratio from growth rate and halve GR (ablation) | RDNet · ImageNet-1K val | top-1 accuracy: 80.4 -> 80.8 | Improved |
| add transition layer every three blocks and raise GR (ablation) | RDNet · ImageNet-1K val | top-1 accuracy: 80.8 -> 82.3 | Improved |
| add patchification stem (ablation) | RDNet · ImageNet-1K val | top-1 accuracy: 82.3 -> 82.4 | No clear effect |
| add channel re-scaling (layer-scale plus squeeze-excitation) (ablation) | RDNet · ImageNet-1K val | top-1 accuracy: 82.6 -> 82.8 | Improved |
| add more activations in the block (ablation) | RDNet-T · ImageNet-1K val | top-1 accuracy: 82.8 -> 81.9 | Hurt |
| use stochastic depth ratio 0.15 (ablation) | RDNet-T · ImageNet-1K val | top-1 accuracy: 81.6 -> 82.8 | Improved |
| use variable growth rate per stage instead of uniform GR 120 (ablation) | RDNet-T · ImageNet-1K val | top-1 accuracy / FLOPs: 83.0 (8.9 GFLOPs) -> 82.8 (5.0 GFLOPs) | Mixed |
| transition ratio 0.5 versus higher ratios (ablation) | RDNet-T · ImageNet-1K val | top-1 accuracy: 0.7: 82.3 -> 0.5: 82.8 | Improved |
| use RDNet-B instead of ConvNeXt-B in CLIP training | CLIP on CC3M/CC12M/RedCaps · ImageNet-1K zero-shot | top-1 accuracy: 51.2 -> 54.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- MogaNet-XL — MogaNet: Multi-order Gated Aggregation Network (ICLR 2024) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Densely Connected Convolutional Networks (CVPR 2017 · similarity 0.94)
- Convolutional Networks with Dense Connectivity (TPAMI 2019 · similarity 0.94)
- Lightweight and Progressively-Scalable Networks for Semantic Segmentation (IJCV 2023 · similarity 0.89)
- DCNAS: Densely Connected Neural Architecture Search for Semantic Image Segmentation (CVPR 2021 · similarity 0.89)
- Vision Transformers for Dense Prediction (ICCV 2021 · similarity 0.88)
- Segmentation-Aware Convolutional Networks Using Local Attention Masks (ICCV 2017 · 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.
