Research map / Label-efficient and robust learning / Self-supervised and contrastive
Rethinking 360deg Image Visual Attention Modelling With Unsupervised Learning.
Yasser Abdelaziz Dahou Djilali, Tarun Krishna, Kevin McGuinness, Noel E. O’Connor
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 self-attention to the contrastive encoder (ResNet) (ablation) | ResNet contrastive encoder + decoder · VR-EyeTracking+Sitzman | KLD: 8.574 -> 7.110 | Improved |
| add self-attention to the contrastive encoder (VGG, lambda 0.7) (ablation) | VGG contrastive encoder + decoder · VR-EyeTracking+Sitzman | KLD: 7.965 -> 7.583 | Improved |
| add self-attention to the contrastive encoder (VGG, lambda 0.5) (ablation) | VGG contrastive encoder + decoder · VR-EyeTracking+Sitzman | KLD: 7.540 -> 7.648 | Hurt |
| add self-attention to the contrastive encoder (VGG, lambda 0.9) (ablation) | VGG contrastive encoder + decoder · VR-EyeTracking+Sitzman | KLD: 7.875 -> 7.648 | Improved |
| add self-attention to the contrastive encoder (ResNet) (ablation) | ResNet contrastive encoder + decoder · Salient360! | KLD: 1.039 -> 0.849 | Improved |
| increase lambda from 0.5 to 0.7 (more weight on local NCE) (ablation) | VGG with self-attention · VR-EyeTracking+Sitzman | KLD: 7.648 -> 7.583 | Improved |
| increase lambda from 0.7 to 0.9 (ablation) | VGG with self-attention · VR-EyeTracking+Sitzman | KLD: 7.583 -> 7.648 | Hurt |
| add sphere-to-plane projections (top-to-front, bottom-to-front) as contrastive views (ablation) | ResNet contrastive encoder · Salient360! | KLD: 0.999 -> 0.849 | Improved |
| add projections as contrastive views (ablation) | ResNet contrastive encoder · Salient360! | CC: 0.479 -> 0.584 | Improved |
| add projections as contrastive views (ablation) | ResNet contrastive encoder · Salient360! | AUC-J: 0.736 -> 0.769 | Improved |
| use frozen unsupervised 360 encoder instead of 2D SalGAN saliency model | saliency prediction · VR-EyeTracking+Sitzman | KLD: 10.195 -> 7.110 | Improved |
| use single ERP stream ResNet model instead of ATSAL | 360 saliency inference · Salient360! 4K images | runtime per image (s): 0.230 -> 0.025 | Improved |
Closest papers by meaning
- Cube Padding for Weakly-Supervised Saliency Prediction in 360° Videos (CVPR 2018 · similarity 0.88)
- Understanding and Visualizing Deep Visual Saliency Models (CVPR 2019 · similarity 0.88)
- Spherical Vision Transformers for Audio-Visual Saliency Prediction in 360$^{\circ }$∘ Videos (TPAMI 2025 · similarity 0.87)
- Spherical DNNs and Their Applications in 360 Images and Videos (TPAMI 2021 · similarity 0.87)
- Learning Spherical Convolution for Fast Features from 360° Imagery (NeurIPS 2017 · similarity 0.87)
- Learning from Unlabelled Videos Using Contrastive Predictive Neural 3D Mapping (ICLR 2020 · similarity 0.87)
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.
