Results and Demos for Small Instance Detection

  • Demo video –  counting and detection results for the UCSD pedestrian dataset (split-’max’), with 1200 frames for testing and 160 frames for training. An example video frame is shown below:

UCSD

  •  F1-score on UCSD Pedestrian, Synthetic Cell and Small Object datasets.  Lambda = 0 means no global count constraints are used.

results_table

 

  • PR curve for `max’ split of UCSD dataset.

ucsd_split01_PR_poster-eps-converted-to

  •  Examples of detection results on the UCSD dataset using various methods based on the object density map.

ucsd_methods

 

  • Results of the proposed detection method on (a) UCSD pedestrians, (b) Synthetic cells, (c) Fish, (d) Honeybees, (e) Flies, and (f) Seagulls. The red dots are the ground truth locations and the green boxes/circles are the detection results. More results here.

examples_all