Datasets:

---

KIT-MRT

/

KITScenes-LongTail

like 11

Follow

Institute of Measurement and Control Systems at KIT 20

Hi, glad you like it!
We are currently targeting June/July.

I see, thank you @jvpc0d3r !

I thought it would be released for the challenge, so I have a follow up if you don't mind:
Is the main goal of the challenge more about evaluating something like dataset transfer capabilities? Or something like achieving the best few-shot setup?
Asking because, from what I understand, 1. in the test set we're measuring trajectory quality as the main metric 2. only a few training samples are available 3. the reasoning trace metric is only a tie-breaker

Thanks in advance for all the clarifications

Exactly, the challenge is about few-shot generalization (or zero-shot), similar to the experiments in our paper (Section 5.2 https://arxiv.org/abs/2603.23607).
Using pre-trained models or pre-training on other datasets is allowed and we encourage that (e.g., we pre-trained the UniAD and DMAD baselines on nuScenes).
Your understanding of the metrics is also correct!

Thank you, a few last questions on the metric if that's ok:

• When the prediction doesn't match any reference trajectory (fourth case), the paper says MMS is 3.5 - CP, but I had understood CP to be calculated by comparing for example the jerk of the prediction with the jerk of its reference trajectory ("Specifically, we consider jerk and tortuosity relative to reference trajectories."). How is it calculated when there's no matched reference traj?
• The lateral and longitudinal thresholds to calculate similarity are said to be based on the Waymo procedure i.e. using the initial speed, but Waymo does this thing where they calculate it at both 3 seconds and 5 seconds (from what I understand they calculate the metric for both waypoints and then average). Is this what you do as well?
• I guess this one is obvious, but just to make sure, if a prediction ends in an overlap region such that it gives similarity = 1 to both expert-like and off-road trajectories for example, you just assign the reference trajectory that has the highest score right?

We are happy that you are competing in this year's challenge!

Regarding your questions:

• In the "unmatched case" w.r.t. the MMS score we still compute the CP relative to the most similar trajectory (low similarity in that case).
• We only compute the metrics for the planning horizon of 5s (not an average of 3s and 5s).
• You are correct, we would give the higher score in that case. Yet, that's unlikely since the "matching areas" are usually smaller than the distance between trajectory endpoints.