eval.md

Evaluate your trained models

In order to conduct correctly the evaluation process, you need to specify:

• The dataset you have trained your model
• The dataset you want to evaluate your model.

The dataset you have trained your model should be specified on the General, Model_type and Dataset sections. If you have trained a model with some specific parameters (heads model, fine tuned, backbone, etc.) you should use the same parameters you have used to train your model.

General parameters for evaluation

Take a look at the Eval section. Here are the details for each variable:

Parameter	Description
eval_on	Name of the datasets to evaluate. Choice between [JAAD, LOOK, PIE]
height	Enable the ablation study on the heights of the pedestrians (see the paper for more details). Choice between [yes, no]
split	Splitting strategy, applicable only if [JAAD] selected above. Choice between [scenes, instances]. Otherwise you can put anything, it will be ignored.
path_data_eval	Path where the built data is stored.

Evaluate on LOOK / Use a trained model on LOOK

If you have trained a model on LOOK and/or you want to evaluate your model on LOOK, you should modify the section LOOK.

Parameter	Description
data	Name of subset to evaluate the model on. Choice between [all, Kitti, JRDB, Nuscenes]
trained_on	Which subset the trained model has been trained. Choice between [all, Kitti, JRDB, Nuscenes]

Evaluate on JAAD or PIE

If you want to evaluate your model on JAAD or PIE, you should modify the JAAD_dataset or PIE_dataset section:

Parameter	Description
path_data	Path where the built data is stored
split	Splitting strategy, applicable only if [JAAD] selected above. Choice between [scenes, instances]. Otherwise you can put anything, it will be ignored.
path_txt	path to the ground truth txt files, this parameter shouldn't be modified if the dataset has been created correctly. Default: ./create_data