Implementation of Social LSTM: Human Trajectory Prediction in Crowded Spaces

Paper

Social LSTM: Human Trajectory Prediction in Crowded Spaces

Dependencies

• torch 1.6.0
• matplotlib 3.2.1
• numpy 1.18.4

Files

There are two models available: SocialLSTM and VanillaLSTM.
Dataset is located in datasets/[dataset name], where each dataset is a collection of training and validating data.
Each file in the dataset is of the form
frame_number pedestrian_number y_coordinates x_coordinates

HOWTO

To train and validate a model against a specific training & validating set, run
python3 main.py mode --dataset [dataset_name] --epoch [epoch_num] --T_obs [observe_step] --T_pred [predict_step]
where mode can be either 's' or 'v'
E.g. to train and validate on "eth" dataset in /datasets, simply run python3 main.py "s" --dataset "eth" --epoch 3
To only validate a chosen model against a validating set, run
python3 main.py mode --dataset [dataset_name] --pure_val_name [model_dir] --T_obs [observe_step] --T_pred [predict_step]
To validate a chosen model against a special validating set, run
python3 main.py mode --special_model [model_dir] --special_file [file_name] --special_start [start_ped] --T_obs [observe_step] --T_pred [predict_step]
Special dataset is the dataset of .pkl file with aligned number of frame numbers. If special dataset is too large to run in one sitting, refer to batchprocess.sh .