Hyunggic !

Bicyclist DPM

This bicycle DPM was designed and trained especially for autototive active safety applications such as PCW (Pedestrian Collision Warning).

We collected bicyclist dataset in various places and weather conditions. It consists of two training sets (set00 ~ set01) and two testing sets (set02 ~ set03), each with 1 or 2 seq files, along with associated annotation files. Due to our purpose, most of samples corresponds to rear view and side view of samples of bicycles. The dataset was very carefully collected based on our investigation about bicycle accident statistics [1]. So, we believe that this dataset could be very valuable for those safety applications.

[1] Bicycle Crash Types: A 1990's Information Guide
W.H. Hunter and W.E. Pein and J.C. Stutts, Federal Highway Administration, 2008

[Typical Samples]

[Typical Samples with annotations]

[Dataset Statistics]

Left pie chart shows the distribution of bicycle samples and right hand figure shows the distribution of bottom lines of all (i.e., 4,284) bounding boxes. Key numbers are:

Bicycle_rear: 2,845
Bicycle_front: 553
Bicycle_right: 129
Bicycle_left: 298

Top row shows a height distribution and bottom row shows ratio distribution of each sub-type.

[Trained Bicyclist Model: Rear (3,398) / Side (429) / Diagonal (459)]

[Result Video]
Bicyclist detection result on Pittsburgh Dataset