INT2 Dataset

INT2

background

Motion forecasting is an important component in autonomous driving systems. One of the most challenging problems in motion forecasting is interactive trajectory prediction, whose goal is to jointly forecasts the future trajectories ofinteracting agents.
The continuously changing map information of the existing dataset makes the definition of the domain challenging. Since INT2 is taken from fixed intersections, peak time and non-peak time data can be used as two distinct domains.
To this end, we present a largescale interactive trajectory prediction dataset named INT2 for INTeractive trajectory prediction at INTersections.

view more +

Rush-hour and Non-rush-hour segments

A comparison between peak time (a, b, c) and trough time (d, e, f) traffic at the same intersection in our INT2 dataset. They are naturally treated as two different trajectory prediction domains.

Information

High quality

INT2 is captured at 16 different intersections by a multi-sensor system and an offline detection and tracking algorithm stack. The multi-sensor system consists of several RGB cameras and LiDARs mounted on poles. Sensors are calibrated routinely. The offline algorithm stack applies state-of-the-art 3D detection and tracking algorithms on raw data without the concern of latency, and fuses the results with production-level rules.
Large scale

INT2 contains about 612,000 scenes with each 1 minute long, counting up to 10,200 hours. By contrast, Waymo Open Motion has about 100,000 scenes with every 20 seconds long, counting up to 570 hours. Apart from the total scale, traffic in each 1-minute segment can cover one complete traffic light cycle.
Rich information

INT2 has rich information in several terms. INT2 provides vectorized road maps at all 16 intersections, which is a common input to state-of-the-art trajectory prediction algorithms. INT2 provides 3D agent boxes for small and large vehicles, pedestrians and cyclists. INT provides temporal traffic light information and links to the lanes they have control over.