search

haomo-ai / MotionSeg3D

[IROS 2022] Efficient Spatial-Temporal Information Fusion for LiDAR-Based 3D Moving Object Segmentation
https://npucvr.github.io/MotionSeg3D/
GNU General Public License v3.0
240 stars 21 forks source link

about evaluation #12

Closed QYChan closed 1 year ago

QYChan commented 1 year ago

Hi, I am trying to reproduce the result in your paper but can not understand the evaluation metric difference between training and evaluation script. In training, I got lI7rSYJ6GS But in evaluating, the result was FZ57xZitix What's differences between two 'iou moving' values? Thanks in advance!

MaxChanger commented 1 year ago

Hi @QYChan, thank you for your interest in our project. A quick answer is that the range image is used during training (without using PointHead). In Salsanext backbone, loss and metrics are based on image-level.(I guess for faster training etc.) In the evaluation, it is consistent with the benchmark, and the metrics are calculated at the point-level. So there will be some difference between the two printed results.