search

exiawsh / StreamPETR

[ICCV 2023] StreamPETR: Exploring Object-Centric Temporal Modeling for Efficient Multi-View 3D Object Detection
Other
563 stars 62 forks source link

about input img_size in Waymo #226

Open YANG-CY-163 opened 2 weeks ago

YANG-CY-163 commented 2 weeks ago

Thank you for such a great work! I have some questions about the setting of the Waymo dataset, what is the image resolution used for training and testing, is it 1280x1920 like the config in Waymo Playgroud?

exiawsh commented 2 weeks ago

We will check the details again. Please wait.

yingfei1016 commented 1 week ago

We use the multiscale training as dict(type='RandomScaleImageMultiViewImage', scales=[0.7, 0.75, 0.8]). So the image size is (1920, 1280) x 0.75 .The code is baed on https://github.com/OpenDriveLab/Birds-eye-view-Perception/blob/master/waymo_playground , but we train the waymo dataset by sliding window manner (due to the ICCV2023 deadline) . The training log and config can be found below: 20230228_065532.log.json petr_waymo_mini_r101_v2_6key_resort_0.75_2grad.txt

YANG-CY-163 commented 1 week ago

I got it. Thank you for your reply! It helps a lot!