search

HanLingsgjk / CSCV

code for ScaleFlow++
BSD 3-Clause "New" or "Revised" License
21 stars 0 forks source link

How to reproduce the results of ScaleRAFT #2

Open letsgofeng opened 1 month ago

letsgofeng commented 1 month ago

Thank you very much for your excellent work, I want to know how to reproduce the results of ScaleRAFT on Scene Flow Evaluation 2015. I used the model you gave (cscv_kitti_42.08.pth) to test and got a worse result (SF-bg: 5.50, SF-fg: 11.71, SF-all: 6.53) than what you submitted (SF-bg: 5.31, SF-fg: 10.18, SF-all: 6.12).

HanLingsgjk commented 4 weeks ago

cscv_kitti-42.08.pth is trained based on KITTI160, and if you want to submit it to the official platform, you should use cscv_kittitest_6.12.pth

letsgofeng commented 4 weeks ago

cscv_kitti-42.08.pth is trained based on KITTI160, and if you want to submit it to the official platform, you should use cscv_kittitest_6.12.pth

Thank you very much for your quick reply. I also tested the cscv_kittitest_6.12.pth model, but still didn’t get the results you submitted. The result I get is SF-bg: 5.33, SF-fg: 11.00, SF-all: 6.27.

More specifically, I comment the “validate_kitti(model)" code and uncomment the “kitti_sceneflow_submission(model.module)” code, then I run the following command: CUDA_VISIBLE_DEVICES=0 python dc_flow_eval.py --model=../CSCV/checkpotins/cscv_kittitest_6.12.pth

HanLingsgjk commented 4 weeks ago

May I ask what depth you are using and if it is Lea_depth

HanLingsgjk commented 4 weeks ago

I have recently made some upgrades to ScaleRAFT, and the latest point count is SF-all 5.82. I will update the code later (within a month). Because the results submitted by KITTI are related to the depth used, I am not sure if you have used the same depth as me.

letsgofeng commented 4 weeks ago

I used ganet, and I checked your code and it also uses ganet. The following are the detailed results: 捕获

This is your result: 捕获

Their D1-bg, D1-fg, and D1-all are the same. Are there any other settings?