search

dtc111111 / Multi-Modal-UAV

We win the 1st place in UG2+, a task in CVPR 2024 UAV Tracking and Pose-Estimation Challenge.
MIT License
21 stars 3 forks source link

How to use MMAUD V1 dataset? (all DSBCAN labels are 0 in my method) #3

Open hxlei0827 opened 3 months ago

hxlei0827 commented 3 months ago

hi, I used "MMAUD V1 Rooftop Simple" from "https://github.com/ntu-aris/MMAUD", and tried to follow your trajectory estimation part. May I know how you deal with the V1 datasets into train and val set for LSTM training? My method is to sample one data from every four data from each type of UAV in V1 dataset to build my val set. However, I got stuck at "building_lidar_360_dataset.py“, because I found all the labels generated from DBSCAN are 0, and the feature set generated from this python file seems correct. May I know if it is due to the inappropriate separation of the train and val dataset? 屏幕截图 2024-08-14 224951 屏幕截图 2024-08-14 225001

hxlei0827 commented 3 months ago

I visualize all the windows(20 frames per window), we can see the uav's trajectory. But why all the labels generate by DBSCAN are 0? 屏幕截图 2024-08-15 163442

ZHOUYI1023 commented 3 months ago

The code logic is simple, just accumulate several frames and use dbscan to cluster the accumulated point clouds. You can visualise the cluster results to check your problem.