DB-3DME: From Dataset to Benchmark for Human-aligned Automatic 3D Mesh Evaluation

Name: DB-3DME: From Dataset to Benchmark for Human-aligned Automatic 3D Mesh Evaluation
Uploaded: 2026-10-07T22:49:49.877641+00:00
Duration: 318 s
Description: DB-3DME created a dataset of 2,619 3D meshes with human ratings to evaluate Vision-Language Models (VLMs) for assessing geometry quality and prompt adherence. The study found that the visual encoder's capability is crucial for accurate evaluation. Fine-tuning an open-source VLM significantly improved performance, setting a new standard for automatic 3D asset assessment.

DB-3DME created a dataset of 2,619 3D meshes with human ratings to evaluate Vision-Language Models (VLMs) for assessing geometry quality and prompt adherence. The study found that the visual encoder's capability is crucial for accurate evaluation. Fine-tuning an open-source VLM significantly improved performance, setting a new standard for automatic 3D asset assessment.

Open the full episode