Research map / Video / Video segmentation and tracking
SDAS: Semantic Data Acquisition System for Minimizing Redundancy and Maximizing Diversity
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SDAS, a markerless motion-capture data acquisition system that uses a semantic similarity metric and Redis-based retrieval to store only non-redundant, diverse motion data, with visualization and LLM-based feature explanation.
Problem
Human motion databases contain redundant identical or similar movements, reducing training efficiency; existing similarity metrics based on joint movements fail to capture semantically similar motions that differ in timing or angle.
Approach
Combine joint-based pose similarity with three semantic features (stability, liveliness, attention) in a weighted similarity metric, and use it in a real-time capture-retrieve-compare pipeline backed by an in-memory database, with LLM-generated textual explanations of semantic features.
Key results
- Semantic similarity metric gives 0.9801 for delayed, 0.8973 for flipped, 0.9448 for extended, and 0.9735 for contracted motions, versus cosine similarity 0.9594, 0.6271, 0.9452, 0.9646.
- Top-1 retrieval accuracy is 1.000 for Walk and Kick and 0.800 to 0.967 for other labels such as Run, Throw, and Punch.
- The reported Top-1 accuracy is above 0.8 for most motions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the Semantic Similarity Metric instead of cosine similarity | Motion similarity evaluation · Delay | similarity: 0.9594 -> 0.9801 | Helped |
| Use the Semantic Similarity Metric instead of cosine similarity | Motion similarity evaluation · Flip | similarity: 0.6271 -> 0.8973 | Helped |
| Use the Semantic Similarity Metric instead of cosine similarity | Motion similarity evaluation · Extend | similarity: 0.9452 -> 0.9448 | Hurt |
| Use the Semantic Similarity Metric instead of cosine similarity | Motion similarity evaluation · Contract | similarity: 0.9646 -> 0.9735 | Helped |
Surprising findings
- Cosine similarity drops sharply (0.6271) on flipped motions, which the authors argue should be judged similar.
Limitations
- Evaluation is small: 10 motions for the metric test and 30 new captures for retrieval.
- Semantic similarity weights w_m and w_f are not specified in the text.
- Only a demonstration system; no quantitative evaluation of redundancy reduction or downstream model training benefit.
Experimental setup
Datasets: CMU Mocap
Benchmarks: CMU Mocap (transformed motions) CMU Mocap retrieval (8 labels, 30 new captures)
Metrics: similarity score Top-1 accuracy
Baselines: cosine similarity Delay Flip Extend Contract
Builds on: SMPL Redis semantic motion features (Lee et al. 2022)
Closest papers by meaning
- Multi-Edge Reinforced Collaborative Data Acquisition for Continuous Video Analytics by Prioritizing Quality over Quantity (AAAI 2025 · similarity 0.80)
- 4DP-QA: Scalable QA for 4D Perception in Vision Language Models (CVPR 2026 · similarity 0.80)
- MotionRAG: Motion Retrieval-Augmented Image-to-Video Generation (NeurIPS 2025 · similarity 0.79)
- Efficient Motion-Aware Video MLLM (CVPR 2025 · similarity 0.79)
- VimoRAG: Video-based Retrieval-augmented 3D Motion Generation for Motion Language Models (NeurIPS 2025 · similarity 0.79)
- Motions as Queries: One-Stage Multi-Person Holistic Human Motion Capture (CVPR 2025 · similarity 0.79)
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.