Research map / 3D vision / Face and shape
LoST: Level of Semantics Tokenization for 3D Shapes
Niladri Shekhar Dutt, Zifan Shi, Paul Guerrero, Chun-Hao Paul Huang, Duygu Ceylan, Niloy J. Mitra, Xuelin Chen
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 RIDA semantic alignment loss (ablation) | LoST (1 token) · Tokenizer reconstruction (1k shapes) | DINO similarity: 0.720 -> 0.731 | Improved |
| add RIDA semantic alignment loss (ablation) | LoST (512 tokens) · Tokenizer reconstruction (1k shapes) | DINO similarity: 0.904 -> 0.921 | Improved |
| add RIDA semantic alignment loss (ablation) | LoST (16 tokens) · Tokenizer reconstruction (1k shapes) | DINO similarity: 0.821 -> 0.814 | Hurt |
| add RIDA semantic alignment loss (ablation) | LoST (1 token) · Tokenizer reconstruction (1k shapes) | DINOv2 similarity: 0.528 -> 0.556 | Improved |
| add RIDA semantic alignment loss (ablation) | LoST (64 tokens) · Tokenizer reconstruction (1k shapes) | DINOv2 similarity: 0.763 -> 0.805 | Improved |
| add RIDA semantic alignment loss (ablation) | LoST · Tokenizer reconstruction (1k shapes) | Chamfer Distance | No clear effect |
| add RIDA semantic alignment loss (ablation) | LoST diffusion decoder · training convergence | training speed: ~40% faster | Improved |
| use RIDA features instead of raw triplane features for retrieval | semantic shape retrieval · Retrieval OOD set | mAP@3: 17.47% -> 28.28% | Improved |
| use RIDA features instead of raw triplane features for retrieval | semantic shape retrieval · Retrieval in-distribution set | mAP@3: 16.42% -> 44.28% | Improved |
| use RIDA instead of direct MSE regression of DINO features (ablation) | semantic extractor · Retrieval in-distribution set | mAP@3: 22.90% -> 44.28% | Improved |
| use RIDA instead of direct MSE regression of DINO features (ablation) | semantic extractor · Retrieval OOD set | Recall@3: 20.63% -> 32.03% | Improved |
| use LoST tokenizer on TRELLIS latents instead of Direct3D triplane latents | LoST (1 token) · Objaverse (128 assets) | DINO similarity: 0.690 -> 0.631 | Hurt |
Closest papers by meaning
- Merge3D: Efficient 3D Multimodal LLMs via Joint 2D-3D Token Merging (CVPR 2026 · similarity 0.90)
- Let Language Constrain Geometry: Vision–Language Models as Semantic and Spatial Critics for 3D Generation (ICML 2026 · similarity 0.89)
- SAR3D: Autoregressive 3D Object Generation and Understanding via Multi-scale 3D VQVAE (CVPR 2025 · similarity 0.88)
- One-shot 3D Object Canonicalization based on Geometric and Semantic Consistency (CVPR 2025 · similarity 0.87)
- G^3-LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding (CVPR 2024 · similarity 0.87)
- LAGD: Local Topological-Alignment and Global Semantic-Deconstruction for Incremental 3D Semantic Segmentation (AAAI 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.
