On Accurate Computation of Trajectory Similarity via Single Image Super-Resolution

---

This repository is the official implementation of On Efficient Computation of Trajectory Similarity via Single Image Super-Resolution, based on the PyTorch.

Requirements

• Ubuntu OS
• PyTorch 1.0+
• Python >= 3.5 (Anaconda3 is recommended)

Preprocessing

---

1. Download dataset

   curl https://archive.ics.uci.edu/ml/machine-learning-databases/00339/train.csv.zip -o data/train.csv.zip
   cd data && unzip train.csv.zip

2. Generate trajectory images

   python3.7 preprocessing.py
   python3.7 traj2img.py

   The hyper parameters used in this experiment are stored in hyper_parameters.json, like the selected city area, image sizes and scale of the cell.

   Note that the generated image data may be too large (200,000 trajectories create image data of nearly 1.3 TB ) to be stored in SSD, and we recommend you to create a soft link to a hard disk drive (HDD), like ln -s PATH_of_your_HDD/image, and store these data in HDD. Or you could set a smaller data amount in traj2img.py(we originally set amount = 200000), like 100000, and accordingly, you also have to change to data size used for training and validation in dataset.py.

   The generated images are saved in image/.

Training

---

python3.7 train.py --cuda --num_epochs 4 --lr 0.0001

This procedure will save the models in directory checkpoint/ at every epoch. We empirically found that 3~4 epochs will produce the model with good enough performance, and longer training process did not bring much improvement during the inference stage.

You can also use the parameter--pretrained PATH_TO_PRETRAINED_MODELto load trained model. We provide our model checkpoint_MyG_3_epoch_4.pt, which is trained after 4 epochs.

Evaluation

---

Generate embedding vectors for trajectory similarity computation:

python3.7 generate_data.py --model_path checkpoint/checkpoint_MyG_3_epoch_4.pt

Compute the trajectory similarity:

python3.7 mutilprocess_test.py --model_path checkpoint/checkpoint_MyG_3_epoch_4.pt