jgnooo/MVADepth

Multi-View-Attention (MVA) Monocular Depth Estimation

Multi-View Attention (MVA) Monocular Depth Estimation

Tensorflow 2 Implementation of MVA monocular depth estimation.
Reference code : DenseDepth

Requirements

• Tensorflow 2.2
• Numpy
• Pilow
• Matplotlib
• Scikit-image 0.15.0
• tqdm

Pre-trained model

• Trained by NYU RGB-D V2 (650 MB)

Dataset

• We use NYU Detph V2 dataset and refer to DenseDepth, See DenseDepth Data section

Usage

• Use pre-trained model

  1. Download pre-trained model weights from above download link above.
  2. Save downloaded model weights to checkpoints directory or your_own_path.

• Use jupyter notebook (example.ipynb)

  1. Make direcotry checkpoints directory.
  2. Locate model weights at checkpoints directory.
  3. Locate the own test images at examples directory.
  4. Go to example.ipynb

Train & Test network

• Train

  • Prepare the dataset for training. (we used the NYU V2 dataset)
  • Run following command.
    python train.py --bs 4 --lr 0.0001 --epochs 20

• Test

  • Prepare your test images in examples directory or your own directory.
  • Make checkpoints directory.
  • Locate model weights at checkpoints directory.
  • Run following command.

  python test.py \
      --model_weights /your/own/path \
      --images_dir /your/own/path \
      --results_dir /your/own/path \
      --gpu your_gpu_number
  

Results