Research map / Detection and segmentation / CNN features
Learning to Exploit Multiple Vision Modalities by Using Grafted Networks
Yuhuang Hu, Tobi Delbruck, Shih-Chii Liu
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 |
|---|---|---|---|
| graft a trained thermal front end instead of feeding thermal frames into the pretrained network | YOLOv3 · FLIR thermal driving | AP50: 39.92 -> 45.27 | Improved |
| use thermal grafted network instead of intensity frames | YOLOv3 · FLIR thermal driving | AP50: 30.36 -> 45.27 | Improved |
| supervised training of 47M parameters instead of grafting 3.2M parameters | Faster R-CNN · FLIR thermal driving | AP50: 45.27 -> 53.97 | Improved |
| use DVS-10 instead of DVS-3 event slices | YOLOv3-GN · MVSEC car detection | AP50: 70.14 -> 70.35 | Improved |
| train on only 40% of the training samples (ablation) | YOLOv3-GN · MVSEC car detection | AP50: 70.35 -> 66.75 | Hurt |
| train on only 10% of the training samples (ablation) | YOLOv3-GN · MVSEC car detection | AP50: 70.35 -> 47.88 | Hurt |
| feed event volume to the pretrained intensity network without a grafted front end (ablation) | YOLOv3-N · MVSEC car detection | AP50: 73.53 -> almost 0 | Hurt |
| combine event-driven and intensity predictions with NMS | YOLOv3 · MVSEC car detection | AP50: 73.53 -> 75.45 | Improved |
| use deeper front end S3 instead of shallower S1 or S2 (ablation) | YOLOv3-GN · FLIR thermal and MVSEC | AP50 | Improved |
| use middle net S5 instead of S4 (ablation) | YOLOv3-GN · FLIR thermal and MVSEC | AP50 | No clear effect |
| add the style loss FSL to FRL and FEL (ablation) | YOLOv3-GN · FLIR thermal and MVSEC | AP50 | No clear effect |
| train front end self-supervised instead of supervised on labeled N-MNIST (ablation) | LeNet · N-MNIST | AP50: 1.11 -> 1.47 | Hurt |
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.
- HATS — HATS: Histograms of Averaged Time Surfaces for Robust Event-Based Object Classification (CVPR 2018) · 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
- Training Object Detectors From Scratch: An Empirical Study in the Era of Vision Transformer (CVPR 2022 · similarity 0.87)
- Pruning Self-Attentions Into Convolutional Layers in Single Path (TPAMI 2024 · similarity 0.85)
- GraftNet: Towards Domain Generalized Stereo Matching With a Broad-Spectrum and Task-Oriented Feature (CVPR 2022 · similarity 0.85)
- Filter Grafting for Deep Neural Networks (CVPR 2020 · similarity 0.85)
- Once-for-All: Train One Network and Specialize it for Efficient Deployment (ICLR 2020 · similarity 0.85)
- Pyramid Grafting Network for One-Stage High Resolution Saliency Detection (CVPR 2022 · 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.
