Research map / Vision-language and multimodal / Multimodal LLMs
SpatialBoost: Enhancing Visual Representation through Language-Guided Reasoning
Byungwoo Jeon, Dongyoung Kim, Huiwon Jang, Insoo Kim, Jinwoo Shin
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 LLM-based decoder supervision instead of SAM decoder (ablation) | DINOv2 ViT-L/14 · ADE20K segmentation | mIoU: 50.1 -> 51.5 | Improved |
| use LLM-based decoder supervision instead of VGGT decoder (ablation) | DINOv2 ViT-L/14 · ADE20K segmentation | mIoU: 45.6 -> 51.5 | Improved |
| use LLM-based decoder supervision instead of linear depth head (ablation) | DINOv2 ViT-L/14 · ImageNet-1K classification | accuracy: 85.7 -> 88.3 | Improved |
| use LLM-based decoder supervision instead of linear segmentation head (ablation) | DINOv2 ViT-L/14 · NYUd depth estimation | RMSE: 0.45 -> 0.32 | Improved |
| use LLM-based decoder instead of SAM decoder (ablation) | DINOv2 ViT-L/14 · ScanQA | BLEU-1: 37.6 -> 40.0 | Improved |
| use forward (pixel->object->scene) turn order instead of reverse (ablation) | SpatialBoost on DINOv2 · ADE20K segmentation | mIoU: 48.4 -> 48.9 | Improved |
| use forward turn order instead of random order (ablation) | SpatialBoost on DINOv2 · NYUd depth estimation | RMSE: 0.36 -> 0.34 | Improved |
| combine 50K single-view and 50K multi-view data instead of 100K single-view only (ablation) | SpatialBoost on DINOv2 · ADE20K segmentation | mIoU: 48.9 -> 49.2 | Improved |
| combine single-view and multi-view data instead of multi-view only (ablation) | SpatialBoost on DINOv2 · NYUd depth estimation | RMSE: 0.36 -> 0.32 | Improved |
| replace simple fine-tuning with SpatialBoost (ablation) | DINOv3 · ADE20K segmentation | mIoU: 56.4 -> 59.7 | Improved |
| replace simple fine-tuning with SpatialBoost (ablation) | DINOv3 · CortexBench robot learning | average score: 75.5 -> 80.8 | Improved |
| replace simple fine-tuning with SpatialBoost (ablation) | OpenCLIP · NYUd depth estimation | RMSE: 0.56 -> 0.40 | Improved |
Closest papers by meaning
- SpatialRGPT: Grounded Spatial Reasoning in Vision-Language Models (NeurIPS 2024 · similarity 0.91)
- From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.91)
- SpatialLadder: Progressive Training for Spatial Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.91)
- SpatiaLab: Can Vision–Language Models Perform Spatial Reasoning in the Wild? (ICLR 2026 · similarity 0.90)
- Stitch and Tell: A Structured Data Augmentation Method for Spatial Understanding (NeurIPS 2025 · similarity 0.89)
- SpatialCLIP: Learning 3D-aware Image Representations from Spatially Discriminative Language (CVPR 2025 · similarity 0.89)
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.
