python train_fisheye.py --batch-size=3 --num-workers=12 --lr=0.001 --load-weight=/home/tian/weights/1221备份/model_sunny_best_model.pth --task-type=sunny

python train_fisheye.py --batch-size=3 --num-workers=12 --lr=0.001 --load-weight=/home/tian/weights/1221备份/model_omnithings_best_model.pth --task-type=omnithings

omnihouse's best model, finetune on our render dataset, test on real scenes.

before ft

Omnidataset's cameras are different (extrincs & intrincs) with ours.

We render dataset with cameras as same as real scenes (FOV220)

Global rmse: 3.0661
Global mae: 2.7942
Global rel_error: 0.8915
Global rmse_0.1_3m: 1.8209
Global mae_0.1_3m: 1.7209
Global rel_error_0.1_3m: 0.8562
Global rmse_0.1_5m: 2.7856
Global mae_0.1_5m: 2.5654
Global rel_error_0.1_5m: 0.8868
Global rmse_0.1_7m: 3.0661
Global mae_0.1_7m: 2.7942
Global rel_error_0.1_7m: 0.8915
Global rmse_0.1_10m: 3.0661
Global mae_0.1_10m: 2.7942
Global rel_error_0.1_10m: 0.8915
Global rmse_0.1_2m: 1.2588
Global mae_0.1_2m: 1.2158
Global rel_error_0.1_2m: 0.7988
Global loss: 144.2654
[EVAL ONLY] RMSE: 3.0661266137522163

python train_fisheye.py --load-weight=tmp/model_omnihouse_best_model.pth --task-type=realscenes --eval-only True --exp-name=test_real_wo_ft

ft on our render dataset (Digital twin camera)

Global rmse: 1.2573
Global mae: 1.0718
Global rel_error: 0.4735
Global rmse_0.1_3m: 0.9502
Global mae_0.1_3m: 0.8142
Global rel_error_0.1_3m: 0.6036
Global rmse_0.1_5m: 1.1482
Global mae_0.1_5m: 0.9830
Global rel_error_0.1_5m: 0.4795
Global rmse_0.1_7m: 1.2573
Global mae_0.1_7m: 1.0718
Global rel_error_0.1_7m: 0.4735
Global rmse_0.1_10m: 1.2573
Global mae_0.1_10m: 1.0718
Global rel_error_0.1_10m: 0.4735
Global rmse_0.1_2m: 0.9325
Global mae_0.1_2m: 0.8220
Global rel_error_0.1_2m: 0.7530
Global loss: 39.7996
[EVAL ONLY] RMSE: 1.2573211640330917

python train_fisheye.py --load-weight=tmp/render119_model_realscenes_last_model.pth --task-type=realscenes --eval-only True --exp-name=test_real_ft_on_my_render