Tensorflow implementation for HIGH-QUALITY-MONOCULAR-DEPTH-ESTIMATION

High Quality Monocular Depth Estimation via a Multi-scale Netowrk and Detail-perserving Objective

1. To trian the model with the full loss of the paper on the NYU Depth V2 dataset,
   python train.py
   --dataset nyu_depth_v2
   --train_file "../filenames/nyu_depth_v2_train_even.txt"
   --val_file "../filenames/nyu_depth_v2_val_2k.txt"
   --cnn_model "resnet_v2_50"
   --decoding_at_image_size
   --output_stride 16
   --multi_grid 1 2 4
   --aspp_rates 4 8 12
   --loss_depth_norm berhu
   --loss_gradient_magnitude_norm l1
   --loss_gradient_magnitude_weight 1.0
   --loss_gradient_direction_norm l1
   --loss_gradient_direction_weight 1.0
   --loss_normal_weight 0.0
   --batch_size 8
   --num_epochs 20
   --learning_rate 1e-4
   --num_gpus 1
   --num_threads 4
   --batch_norm_epsilon 1e-5
   --batch_norm_decay 0.9997
   --l2_regularizer 1e-4

2. To evaluate on the model,
   python test_nyu_depth_v2.py --process_id_for_evaluation pid(12582)
   You can download the trained model and results of the testset from Badidu NetDisk,
   link: https://pan.baidu.com/s/1tihXtR72Y-M_PyyOhYEryQ code: gmfx

Test performance:
|abs_rel: 0.127 |sq_rel: 0.088 |rmse: 0.468 |rmse_log: 0.165 |log10: 0.054 |acc1: 0.841 |acc2: 0.967 |acc3: 0.993