search

mileyan / pseudo_lidar

(CVPR 2019) Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving
https://mileyan.github.io/pseudo_lidar/
MIT License
975 stars 216 forks source link

Disparity results. #10

Open LifeBeyondExpectations opened 5 years ago

LifeBeyondExpectations commented 5 years ago

Thanks for sharing this nice work with code. Can you share the results of the disparity / depth accuracy from your newly modified PSMnet (using KITTI Stereo Benchmark)?? I download your code and achieve the metrics but it is too bad so I am not sure whether the pre-trained weight is correct.

So far, the disparity and depth metric that I obtained are below, 2

1

mileyan commented 5 years ago

Hi Jaesung,

The released model is trained on 3,712 KITTI 3D object detection training images. And the disparity ground truth is generated by ourselves, which might be slight different with they way used on kitti stereo dataset. You can evaluate it by doing the 3d object detection evaluation.

Best

xiazhiyi99 commented 4 years ago

Hi Jaesung,

The released model is trained on 3,712 KITTI 3D object detection training images. And the disparity ground truth is generated by ourselves, which might be slight different with they way used on kitti stereo dataset. You can evaluate it by doing the 3d object detection evaluation.

Best

Hello mileyan, thanks for sharing your working! I have problems about disparity result too. When I tried to run the code, I got a terrible result. It turned out to be the PSMnet's problem. So I downloaded your finetuned model, but I got sparse result like this image running on 2 1080ti, and here is the command:

python ./psmnet/submission.py \
    --loadmodel ./psmnet/kitti_3d/finetune_300.tar \
    --datapath ./KITTI/object/testing/ \
    --save_path ./KITTI/object/testing/predict_disparity

It doesnt seem correct, but I really dont know where I got wrong. I really appreciate it if you could provide some advise. Forgive my naive.

sarimmehdi commented 4 years ago

Hello. I am getting same results as above with the pre-trained model. The depth image looks like a mess. I am running it on new images however but the results are still quite bad. Is this normal?