search

TRI-ML / packnet-sfm

TRI-ML Monocular Depth Estimation Repository
https://tri-ml.github.io/packnet-sfm/
MIT License
1.23k stars 242 forks source link

Unable to reproduce the results of PackNet-SAN on DDAD #167

Open zhangjiannan1 opened 3 years ago

zhangjiannan1 commented 3 years ago

Here are the results of my training on DDAD： || | E: 32 BS: 2 - SemiSupCompletionModel LR (Adam): Depth 5.00e-05 | || | METRIC | abs_rel | sqr_rel | rmse | rmse_log | a1 | a2 | a3 | |***| | * /DDAD/ddad_train_val/ddad.json/val (camera_01:) | || | DEPTH | 0.155 | 3.528 | 14.022 | 0.240 | 0.804 | 0.924 | 0.967 | | DEPTH_PP | 0.154 | 3.425 | 13.910 | 0.238 | 0.805 | 0.925 | 0.967 | | DEPTH_GT | 0.148 | 3.050 | 13.874 | 0.235 | 0.812 | 0.926 | 0.969 | | DEPTH_PP_GT | 0.147 | 2.972 | 13.770 | 0.233 | 0.814 | 0.927 | 0.969 | ||

Can you provide some advice or help，thanks！

VitorGuizilini-TRI commented 3 years ago

Are these depth predicted or completed results? Can you please provide both?

zhangjiannan1 commented 3 years ago

This is the result of deep completion：

| * /DDAD/ddad_train_val/ddad.json/val (camera_01:lidar) | |***| || | DEPTH | 0.223 | 4.379 | 15.935 | 0.294 | 0.728 | 0.883 | 0.947 | | DEPTH_PP | 0.217 | 4.071 | 15.747 | 0.288 | 0.734 | 0.884 | 0.948 | | DEPTH_GT | 0.195 | 3.952 | 17.328 | 0.297 | 0.714 | 0.880 | 0.944 | | DEPTH_PP_GT | 0.191 | 3.759 | 17.192 | 0.292 | 0.718 | 0.883 | 0.946 | ||

The strange thing is that the result of completion is worse than predicted.

Note: I only trained "camera_01"，and I add the “input_depth_type: ['lidar']” in “datasets:train:” . I’m not sure if this must be added, but I noticed this attribute in kitti’s training profile.

here is my dataset config: datasets: augmentation: image_shape: (384, 640) train: batch_size: 2 num_workers: 16 dataset: ['DGP'] path: ['/DDAD/ddad_train_val/ddad.json'] split: ['train'] depth_type: ['lidar'] input_depth_type: ['lidar'] cameras: [['camera_01']] repeat: [1] validation: num_workers: 8 dataset: ['DGP'] path: ['/DDAD/ddad_train_val/ddad.json'] split: ['val'] depth_type: ['lidar'] input_depth_type: [''] cameras: [['camera_01']] test: num_workers: 8 dataset: ['DGP'] path: ['/DDAD/ddad_train_val/ddad.json'] split: ['val'] depth_type: ['lidar'] input_depth_type: ['','lidar'] cameras: [['camera_01'],['camera_01']]

RobinhoodKi commented 3 years ago

Do you guys tried to train PackNet-sfm from scratch ? I tried three times but still can not reproduce the results especially abs_rel to 0.173 . I can only obtain value between 0.2 and 0.21 . Can you guys reproduce it ? @ @zhangjiannan1

zhangjiannan1 commented 3 years ago

Do you guys tried to train PackNet-sfm from scratch ? I tried three times but still can not reproduce the results especially abs_rel to 0.173 . I can only obtain value between 0.2 and 0.21 . Can you guys reproduce it ? @ @zhangjiannan1

I haven't tried it, but have you tried adjusting the learning rate？

zhangjiannan1 commented 3 years ago

15616462084436480_rgb 15616462084436480_viz This is my result, look forward to your help.

baiyancheng20 commented 2 years ago

Here are the results of my training on DDAD： |****| | E: 32 BS: 2 - SemiSupCompletionModel LR (Adam): Depth 5.00e-05 | |****| | METRIC | abs_rel | sqr_rel | rmse | rmselog | a1 | a2 | a3 | |****| | * /DDAD/ddad_train_val/ddad.json/val (camera01:) | |****| | DEPTH | 0.155 | 3.528 | 14.022 | 0.240 | 0.804 | 0.924 | 0.967 | | DEPTH_PP | 0.154 | 3.425 | 13.910 | 0.238 | 0.805 | 0.925 | 0.967 | | DEPTH_GT | 0.148 | 3.050 | 13.874 | 0.235 | 0.812 | 0.926 | 0.969 | | DEPTH_PP_GT | 0.147 | 2.972 | 13.770 | 0.233 | 0.814 | 0.927 | 0.969 | |***|

Can you provide some advice or help，thanks！

I got almost the same results. Have you fixed the problem?