struct2depth_eval

This code is modified from "struct2depth" to use and evaluate depth_from_video_in_the_wild.

Abs Rel Error calculation code is from SfM Learner.

Example input

Train example

• Data Preparation
• ImageNet Checkpoint Preparation

nohup python -m depth_from_video_in_the_wild.train \
--data_dir /home/ubuntu/data/kitti_result_all_20200715 \
--checkpoint_dir=/home/ubuntu/data/kitti_experiment_checkpoint_20200716 \
--imagenet_ckpt=/home/ubuntu/data/ResNet18/model.ckpt \
--train_steps=1000000 &

Getting Abs Rel Error

python kitti_eval/eval_depth.py --kitti_dir=/home/ubuntu/data/raw_data_KITTI/ --pred_file=/home/ubuntu/data/result_20200716_279296/result.npy

Finetuning with the video taken in Saitama

nohup python -m depth_from_video_in_the_wild.train \
--data_dir /home/ubuntu/Sayama/out \
--checkpoint_dir=/home/ubuntu/data/kitti_experiment_checkpoint_20200716 \
--imagenet_ckpt=/home/ubuntu/data/ResNet18/model.ckpt \
--train_steps=1000000 &

Evaluation

Before fine tuning

Getting Predicted Depth

python inference_dfv.py \
    --logtostderr \
    --file_extension png \
    --depth \
    --egomotion false \
    --input_dir /home/ubuntu/Sayama/tmpdir/2020_08_04/video1top_png/image_02/data/ \
    --output_dir /home/ubuntu/Sayama/result_video1top_273486/ \
    --model_ckpt /home/ubuntu/data/kitti_experiment_checkpoint_20200716/model-273486

After fine tuning

Getting Predicted Depth

python inference_dfv.py \
    --logtostderr \
    --file_extension png \
    --depth \
    --egomotion false \
    --input_dir /home/ubuntu/Sayama/tmpdir/2020_08_04/video1top_png/image_02/data/ \
    --output_dir /home/ubuntu/Sayama/result_video1top_279296/ \
    --model_ckpt /home/ubuntu/data/kitti_experiment_checkpoint_20200716/model-279296