Research map / Detection and segmentation / CNN features
Unsupervised Dense Prediction using Differentiable Normalized Cuts
Yanbin Liu, Stephen Gould
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 structural regularization with horizontal, vertical and uniform mask templates (ablation) | Differentiable NCut self-supervised dense learning · DUT-OMRON saliency, Simple post-processing | IoU: 48.8 -> 56.4 | Improved |
| add the projector g (ablation) | Differentiable NCut self-supervised dense learning · DUT-OMRON saliency, Simple post-processing | IoU: 52.2 -> 56.4 | Improved |
| use NCut mask consistency loss instead of pixel-level consistency (ablation) | Self-supervised dense learning · DUT-OMRON saliency, Simple post-processing | IoU: 46.0 -> 56.4 | Improved |
| replace frozen DINO with the NCut-finetuned backbone in TokenCut | TokenCut · DUT-OMRON saliency | IoU: 53.3 -> 55.7 | Improved |
| replace DINO backbone with finetuned backbone in MOVE and apply bilateral solver | MOVE · DUT-OMRON and DUTS-TE saliency with bilateral solver | IoU: DUT-OMRON 63.6 -> 63.0; DUTS-TE 68.7 -> 66.9; maxFbeta 73.4 -> 74.7 and 82.1 -> 83.7 | Mixed |
| apply finetuned backbone to object discovery post-processing methods | TokenCut, FOUND, MOST · VOC07 object discovery | CorLoc: 68.8 -> 71.3 (TokenCut); 72.5 -> 73.5 (FOUND); 74.8 -> 75.6 (MOST) | Improved |
| finetune MAE and MoCoV3 with the framework for two epochs | MAE ViT-B/16, MoCoV3 ViT-S/16 · Saliency, Simple post-processing | average IoU: +8.7 (MAE), +9.9 (MoCoV3) | Improved |
| finetune ResNet50 with the framework | DINO ResNet50 · Saliency, Simple post-processing | average IoU: +6.4 | Improved |
| replace DINO backbone with finetuned backbone for unsupervised semantic segmentation | ViT-S/16 on Pascal VOC12 · K=500 and K=21 | mIoU: 17.4 -> 19.0 (K=500); 4.6 -> 5.8 (K=21) | Improved |
| finetune FOUND/MOVE backbones with the framework on ViT-S/8 (small saliency gains) | MOVE · ECSSD saliency | IoU: 83.0 -> 83.1 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- MOVE — MOVE: Unsupervised Movable Object Segmentation and Detection (NeurIPS 2022) · better on 3 of 5 shared results
- TokenCut — TokenCut: Segmenting Objects in Images and Videos With Self-Supervised Transformer and Normalized Cut (TPAMI 2023) · better on 4 of 4 shared results
- MOST — MOST: A Multi-Oriented Scene Text Detector With Localization Refinement (CVPR 2021) · better on 3 of 3 shared results
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
- Vision Transformers: From Semantic Segmentation to Dense Prediction (IJCV 2024 · similarity 0.86)
- AiluRus: A Scalable ViT Framework for Dense Prediction (NeurIPS 2023 · similarity 0.86)
- Expediting Large-Scale Vision Transformer for Dense Prediction without Fine-tuning (NeurIPS 2022 · similarity 0.85)
- Self-Supervised Transformers for Unsupervised Object Discovery Using Normalized Cut (CVPR 2022 · similarity 0.85)
- ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions (CVPR 2024 · similarity 0.85)
- Memory Efficient Transformer Adapter for Dense Predictions (ICLR 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.
