RGB-D Domain Adaptation

Machine and Deep Learning Project @ Politecnico di Torino, Italy

The purpose, the achieved results and the theory behind these experiments are thoroughly explained in the project report.

1. Description
2. Requirements
3. Usage
4. References

Description

The main components of the code are the following:

├── main.py: script that can be used to run the main experiments
├── modules
│   ├── datasets.py: dataset handling classes
│   ├── net.py: implementation of the PyTorch module neural networks
│   ├── training_methods*.py: set of helper methods including all possible variations of the training procedure
│   └── transforms.py: custom classes for image transformation
├── python_scripts: set of scripts used for all the experiments
└── results: complete set of achieved results

Requirements

Please check the requirements before running the experiments. You can find the necessary packages in the requirement.txt file. The Python version used to carry out all the experiments is 3.6.9.

Usage

To run one of the experiments (between DA with relative rotation, Stepwise AFN and the combination of the two) just run

python3 main.py

The RR experiment will run with default parameters, if desired, provide options to the command according to the help description:

usage: main.py [-h] [--data_root DATA_ROOT] [--ram] [--no-ram]
               [--ckpt_dir CKPT_DIR] [--result_dir RESULT_DIR]
               [--batch_size BATCH_SIZE] [--epochs EPOCHS] [--lr LR]
               [--class_num CLASS_NUM] [--dropout_p DROPOUT_P]
               [--momentum MOMENTUM] [--weight_decay WEIGHT_DECAY]
               [--step_size STEP_SIZE] [--gamma GAMMA] [--lambda LAMDA]
               [--entropy_weight ENTROPY_WEIGHT] [--dr DR] [--radius RADIUS]
               [--weight_L2norm WEIGHT_L2NORM]
               [--experiment {rr,safn,safn-rr,hafn,hafn-rr}]

optional arguments:
  -h, --help            show this help message and exit
  --data_root DATA_ROOT
  --ram                 dataset stored in main memory
  --no-ram              dataset not stored in main memory, slower processing
  --ckpt_dir CKPT_DIR   leave default if checkpoint not desired
  --result_dir RESULT_DIR
  --batch_size BATCH_SIZE
  --epochs EPOCHS
  --lr LR
  --class_num CLASS_NUM
  --dropout_p DROPOUT_P
  --momentum MOMENTUM
  --weight_decay WEIGHT_DECAY
                        gamma factor for StepLR scheduler
  --step_size STEP_SIZE
                        step size for StepLR scheduler
  --gamma GAMMA
  --lambda LAMDA        weight for pretext loss
  --entropy_weight ENTROPY_WEIGHT
                        weight for entropy loss
  --dr DR               step size for SAFN
  --radius RADIUS       shared fixed R value for HAFN
  --weight_L2norm WEIGHT_L2NORM
                        weight of AFN loss
  --experiment {rr,safn,safn-rr,hafn,hafn-rr}
                        select the experiment to run:`rr`: domain adaptation
                        with relative rotation,`safn`: hard afn e2e,`safn-rr`:
                        stepwise afn and relative rotation DA,`hafn`: hard afn
                        e2e,`hafn-rr`: hard afn and relative rotation DA

References

[1] M. R. Loghmani, L. Robbiano, M. Planamente, K. Park,B. Caputo, and M. Vincze. Unsupervised domain adaptation through inter-modal rotation for rgb-d object recognition,2020

[2] R. Xu, G. Li, J. Yang, and L. Lin. Larger norm more transferable: An adaptive feature norm approach for unsuperviseddomain adaptation, 2018