What's the meaning of the attributes in the result json file?

[Kevin6liu6]

I was able to generate adversarial scenarios following the instructions, and got several BEV images/videos along with a json file. I printed the keys and got ['N', 'dt', 'map', 'lw', 'sem', 'past', 'fut_init', 'fut_adv', 'fut_internal_ego', 'fut_sol', 'attack_agt', 'attack_t', 'z_adv', 'z_sol', 'z_prior']. I am wondering what are the meanings of these keys.

[davrempe]

What you probably care most about is past and fut_adv which are the past trajectories and future trajectories for all agents in the final adversarial scenario.

A full explanation of all fields is as follows:

• N- the number of agents
• dt - timestep size of scenario
• map - name of the map in the dataset that this scenario is on
• lw - the length and width of the bounding boxes for each agent
• sem - the semantic category of each agent indicated by a one-hot vector
• past - the past trajectories of agents in the scene
• fut_init - the future trajectories of agents used to initialize the optimization
• fut_adv - the future trajectories of agents in the final adversarial scenario (i.e. the output of the adversarial optimization)
• fut_internal_ego - this is our method's prediction of the ego agent trajectory after adversarial optimization
• fut_sol- the solution trajectory for the ego to avoid the adversarial attack (i.e. the output of the solution optimization)
• attack_agt - the index of the "attacking" agent, i.e. the one that causes the collision or gets closest
• attack_t - the timestep of the attack (collision) or the timestep that's closest to a collision
• z_adv - the VAE latent vector for the final adversarial scenario
• z_sol - the VAE latent vector for the final solution scenario
• z_prior - the mean and variance of the latent distribution output by the encoder of our model

The evaluation script shows how to read in these json files and visualize them.