Research map / Architectures and efficiency / CNN design
Depth CNNs for RGB-D Scene Recognition: Learning from Scratch Better than Transferring from RGB-CNNs
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 |
|---|---|---|---|
| Replace transferred Places-CNN depth features with D-CNN features | Depth recognition · SUN RGB-D | accuracy (%): 38.9 -> 42.4 | Helped |
| Use D-CNN depth features instead of FT-Places-CNN depth features in fusion | RGB-D-CNN · SUN RGB-D | accuracy (%): 46.9 -> 52.4 | Helped |
| Replace FT-Places-CNN depth features with D-CNN features in RGB-D fusion | RGB-D-CNN · NYUD2 | accuracy (%): 59.5 -> 65.8 | Helped |
| Train the depth network with weakly supervised patches rather than Alex-CNN from scratch on full images (ablation) | Alex-CNN · SUN RGB-D depth features at fc8 | accuracy (%): 32.8 -> 37.5 | Helped |
| Use the reduced D-CNN architecture rather than Alex-CNN with weakly supervised patch training (ablation) | Depth CNN · SUN RGB-D depth features at fc8 | accuracy (%): 37.5 -> 41.2 | Helped |
| Train bottom layers from scratch rather than transfer Places-CNN weights (ablation) | Alex-CNN · SUN RGB-D depth features at pool1 | accuracy (%): 17.2 -> 22.3 | Helped |
| Use weakly supervised patch training rather than training Alex-CNN from scratch (ablation) | Alex-CNN · SUN RGB-D depth features at conv3 | accuracy (%): 29.8 -> 35.1 | Helped |
| Fine-tune only the top layers of Places-CNN rather than adapting bottom layers (ablation) | Places-CNN depth transfer · SUN RGB-D | accuracy (%): The paper reports 36.5% for FT-bottom and says FT-top does not help significantly until bottom convolutional layers are | Hurt |
| Use weighted SVM instead of ordinary SVM | D-CNN · SUN RGB-D depth | accuracy (%): 41.2 -> 42.4 | Helped |
Closest papers by meaning
- Discriminative Multi-Modal Feature Fusion for RGBD Indoor Scene Recognition (CVPR 2016 · similarity 0.91)
- Modality and Component Aware Feature Fusion For RGB-D Scene Classification (CVPR 2016 · similarity 0.90)
- RGB-D Salient Object Detection via 3D Convolutional Neural Networks (AAAI 2021 · similarity 0.89)
- RGBD Based Dimensional Decomposition Residual Network for 3D Semantic Scene Completion (CVPR 2019 · similarity 0.89)
- DFormer: Rethinking RGBD Representation Learning for Semantic Segmentation (ICLR 2024 · similarity 0.88)
- DFormerv2: Geometry Self-Attention for RGBD Semantic Segmentation (CVPR 2025 · 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.