While preparing our submission, we overlaid the provided test masks onto the corresponding test views and noticed that the masked region covers roughly 9–28% of the frame across most sequences. If we understand the evaluation correctly, this means metrics are computed over that fraction of each target view. We wanted to check whether this is the intended behaviour — for example, restricting evaluation to a specific anatomical or tool-free region — or whether we may be applying the masks incorrectly on our side. Attached is a sheet showing the overlay for several sequences. Happy to provide more detail if useful. Thanks! ${imageLink?synapseId=syn76599289&align=None&scale=70&responsive=true&altText=}

Created by Gabriel Perez g.perezsantamaria
Hi, the corrected metrics.py is included in the latest NVS baseline Docker image: docker.synapse.org/syn74277461/imed-nvs-baseline:v1 It is also available on GitHub here: https://github.com/smbonilla/Endo-4DGS/blob/docker_compatible/metrics.py Best regards, Tianyi
Hi, Are there any news on the correct metrics.py? Is it going to be released?
Hello,  Thank you for checking this.  We identified an issue in the organizer-side reprojection mask generation. The depth maps are provided at 512×640 resolution, while the intrinsics in K.txt correspond to the 1024×1280 RGB images. The source-camera intrinsics were not rescaled before depth unprojection, which made the reprojection support masks artificially small. Please continue to submit full-resolution RGB renders. There is no need to crop or pre-mask your predictions. We are updating and validating the organizer-side evaluator and will rescore any affected public-validation submissions. We will post another update once this has been completed.  We apologize for the earlier inaccurate explanation, and thank you again for bringing this to our attention.  Best regards,  Thank you for raising this.

NVS task - low test mask coverage of image, is this intended? page is loading…