Here we provide pre-computed detections with segmentation masks for KITTI MOTS and MOTSChallenge.

For tools and examples please visit https://github.com/VisualComputingInstitute/mots_trackingonly_tools
The project page can be found at https://www.vision.rwth-aachen.de/mots
For questions please contact Paul Voigtlaender via voigtlaender@vision.rwth-aachen.de

These detections have to be used for the Tracking Only Challenge (task 3) of the MOTS Workshop Competiton 2020 (https://motchallenge.net/workshops/bmtt2020/tracking.html).
Note that we provide detections for all sequences of the full KITTI Tracking and MOTChallenge datasets, so there might be detections for some extra sequences which are not (yet) annotated on a pixel level and/or which will not be used for the challenge.

The detection format for each line is the following:
time_frame bbox_x0 bbox_y0 bbox_w bbox_h score class_id img_height img_width rle

An example line:
3 0.42776060104370117 194.42828369140625 11.944610118865967 41.77899169921875 0.2072293758392334 2 375 1242 U6W1_:2O01O000O3M4L2M3M4L7IaWR>

which means
time frame: 3
bounding box (x0, y0, w, h): (0.42776060104370117, 194.42828369140625, 11.944610118865967, 41.77899169921875)
detection confidence score (from 0 to 1): 0.2072293758392334
class id (see below): 2
image height: 375
image width: 1242
rle (run-length encoded mask, like in coco): U6W1_:2O01O000O3M4L2M3M4L7IaWR>

The class ids are the following:
car 1
pedestrian 2

To decode rles, you can use the cocotools from https://github.com/cocodataset/cocoapi

The easiest way to load the detections is using our provided tools from https://github.com/VisualComputingInstitute/mots_trackingonly_tools
