DADA: Dual-Alignment Domain Adaptation for Pedestrian Trajectory Prediction

Setup

Environment
All models were trained and tested on Ubuntu 18.04 with Python 3.8.17 and PyTorch 2.0.1 with CUDA 11.8. You can start by creating a conda virtual environment:

conda create -n dada python=3.8 -y
source activate dada
pip install -r requirements.txt

Dataset
Preprocessed ETH and UCY datasets are included in this repository, under ./datasets/.
Among these datasets, train_origin, val and test are obtained directly from the ETH and UCY datasets, and train is obtained after the DLA processing.
We have provided an example A2B in ./datasets/A2B/, which demonstrates how to set up the dataset for a particular cross-domain task. If you want to construct another S2T dataset, please follow the step below (here use B2C dataset as an example):

create a folder named B2C under ./datasets/;
create four folders named train_origin, train, val and test under ./datasets/B2C/;
put the B-domain (HOTEL) training set into new-created train_origin and train folders; put the C-domain (UNIV) validation set into val folder; put the C-domain testing set into test folder;
train the corresponding DLA model to automatically generate the aligned source data in train folder.

Baseline Models
This repository supports three baseline models: Social-GAN, Trajectron++ and TUTR. Their DADA-modified source code are in ./models/.

Quick Start

To train and evaluate our DADA-model on the A2B task at once, we provide a bash script train.sh for a simplified execution.

bash ./train_DADA.sh -b <baseline_model>  # quickly train
bash ./test_DADA.sh -b <baseline_model>  # quickly evaluate

where <baseline_model> could be sgan, trajectron++ or tutr.
For example:

bash ./train_DADA.sh -b sgan  # quickly train
bash ./test_DADA.sh -b sgan  # quickly evaluate

Detailed Training

Training for DLA

The DLA network could to be trained by:

cd ./DLA/
python train_DLA.py  --subset <task_S2T>

For example:

python train_DLA.py  --subset A2B

After finishing training, the aligned source data will be automatically generated in ./datasets/subset/train/.

Training for Prediction Models

Given that our repository supports three baseline models, here we take the Social-GAN as example.

Training for Baseline
The baseline model is directly trained without DLA data:

cd ./models/sgan/scripts/
python train.py --dataset_name <task_S2T>

Training for DLA
The DLA model is trained with DLA data, so you just need to modify the train_set path to f'../../../datasets/{args.dataset_name}/train' and modify the checkpoint_save path to '../checkpoint/checkpoint_DLA'.

Training for DADA
The DADA model is further embedded an discriminator w.r.t DLA model during training phase:

cd ./models/sgan/scripts/
python train_DADA.py --dataset_name <task_S2T>

You can find the code of the discriminator structure and its training procedure.

Detailed Evaluation

Given that our repository supports three baseline models, here we take the Social-GAN as example.

Pretrained Models
We have included pre-trained models in ./models/sgan/checkpoint/ folder that can be directly used to evaluate models.

You can simply view the DADA evaluation result for A2B task by running:

cd ./models/sgan/scripts/
python evaluate_model.py --dataset_name <task_S2T>

To view the baseline and DLA evaluation result, you just need to modify the checkpoint_load path.

jackyzengl/DADA

DADA: Dual-Alignment Domain Adaptation for Pedestrian Trajectory Prediction

Setup

Quick Start

Detailed Training

Training for DLA

Training for Prediction Models

Detailed Evaluation

Experimental results

ADE results

FDE results

Visualization