What processing has been done on the tracking dataset

[ffiioonnaa]

Hi, thanks for your word! I have the following questions:

1. Do these coordinates correspond to first 32 second at 1fps?
2. Are the bounding boxes in absolute coordinates, and Does the length greater than 1000 need to be normalized to 1000?
3. Is there a situation where, due to the overall token limit, the model has scaled the video but the bounding boxes have not been scaled? How is this situation generally handled?

[tulerfeng]

Hi, thanks for your question.

1. Yes, we use 32 frames (including the first frame) at 1fps for training and evaluation, which follows "VideoChat-R1: Enhancing Spatio-Temporal Perception via Reinforcement Fine-Tuning". However, VideoChat-R1 only use 8 frames for evaluation while we use 32 frames, which is harder than them.

2. No, they are relative coordinates, normalized to the range [0, 1000] according to the resolution, regardless of whether the original dimension exceeds 1000.

This design follows the coordinate convention used in Qwen2-VL and Qwen3-VL, which adopt 0–1000 relative coordinates. In contrast, only Qwen2.5-VL uses absolute coordinates.

Moreover, most of commercial models, such as Gemini 2.5 Pro and Seed 1.5-VL, also employ the 0–1000 relative coordinate scheme. Therefore, to maintain consistency with these models, we adopt the same coordinate representation in both the CoT reasoning and the final answers.

3. No, since we use relative coordinates, this issue does not exist.