How to Evaluate Qwen3-VL on the TraceSpatial Benchmark?

[k00dj-19]

Thanks for the great work.

I’m trying to reproduce your reported results by running Qwen3-VL inference on the TraceSpatial benchmark. However, when I use the system prompt provided in your repository, the output format differs from what the evaluation script expects, which results in a metric score of 0.0 or scores that do not match those reported in the paper.

Could you please share the exact prompting or evaluation setup you used (e.g., system prompt, output format constraints, or post-processing steps) to obtain the reported results?

[Zhoues]

@k00dj-19

Thank you very much for your attention and liking of RoboTracer. We have fixed the bug in the prompt of Evaluation and updated how to use Qwen3VL for evaluation. For details, please refer to README.md.

[k00dj-19]

@Zhoues
Thank you for the update. However, even when running inference following your instructions, the 3D evaluation results are still lower than the performance reported in the paper. Could you please let me know how you achieved those results?

[fanzhenxuan]

same question

[godnpeter]

Me too

[Hyperwjf]

same question