Request for original videos without annotation overlays for aerial–ground VLM research
Dear CVUTD authors,
Thank you for creating and sharing the CrossViewUrbanTrafficDataset. We greatly appreciate the effort involved in collecting aerial and ground footage and providing the accompanying annotations. Your dataset offers a valuable resource for studying cross-view perception in real-world urban environments.
We are developing a benchmark for evaluating spatial reasoning in vision-language models using paired aerial and ground views. Our tasks include cross-view object matching, counting, spatial relationships, and viewpoint selection, and CVUTD is particularly relevant to this work.
After inspecting the released videos, we found versions without annotation overlays for the first two scenes, while the other scene videos we located contain bounding boxes, object IDs, class labels, and other rendered markers. These overlays present two difficulties for our evaluation: they can reveal information that models should infer from the visual scene, and they can obscure object details needed for reliable cross-view reasoning.
We have explored interpolation and video inpainting to remove the overlays. However, the results still contain residual markings or reconstructed details whose accuracy we cannot verify. Since our benchmark depends on preserving object identity and spatial evidence, access to the original footage would be substantially more reliable.
Would it be possible to share the original aerial and ground videos before the annotations were rendered, particularly for the remaining scenes? Ideally, we would like to retain their correspondence with the released annotations. If these videos are already available elsewhere, we would be grateful for directions to the appropriate files.
We fully understand that sharing may be subject to privacy, licensing, storage, or other restrictions. We would be happy to follow your access requirements, discuss permitted benchmark use, and provide proper citation and attribution.
Thank you for your time, consideration, and contribution to the research community.
Best regards,
ShengHong Yi