Menu

Multiple Object Tracking

In the RAIL-BENCH Multiple Object Tracking challenge, the task is to track all persons across consecutive video frames. All scenes are caputured from a front-facing camera mounted on a train, while it approaches a crowded train station and passes through it.

Dataset

The RAIL-BENCH Tracking dataset consists of 4 video sequences containing between 179 and 249 frames and recorded at 10 fps. In contrast to the other RAIL-BENCH challenges, the entire dataset is treated as the test set. The four sequences are available for download:

Annotation Policy

  • Each person is annotated with an axis-aligned bounding box with a unique ID across consecutive frames to capture temporal consistency.
  • The occlusion level of each bounding box is indicated with 0 (0-24 %), 1 (25-49 %), 2 (50-74 %), 3 (75-99 %), or 4 (100 %).
  • Annotations are terminated for objects that permanently leave the field of view. Objects that reappear — even after being fully occluded for several frames — retain their original ID and are continuously labeled.
  • As done for the RAIL-BENCH Object dataset, large crowds of people that cannot be clearly separated into individual instances are annotated with a single bounding box and a special "crowd" flag.

RAIL-BENCH Tracking Challenge

In the RAIL-BENCH Tracking challenge, predictions on the test set are evaluated using the higher order tracking accuracy (HOTA).

How to Participate

  1. Train your model with any data sources, for instance the RAIL-BENCH Object dataset.
  2. Download the RAIL-BENCH Tracking dataset and run inference on the test set.
  3. Soon, we will publish the official CodaBench challenge, where you can submit your predictions and get evaluated on the hidden ground truth annotations.