vaccovecrana / kimaris

Java port for Pixel Intensity Comparison-based Object detection

kimaris

kimaris is a Java implementation of Pixel Intensity Comparison-based Object detection.

Requires Java 11 or later.

Reference code is ported from pico.

Installation

Available in Maven Central.

implementation("io.vacco.kimaris:kimaris:<LATEST_VERSION>")

This library is able to evaluate cascades in real time, as well as training custom cascades with training datasets which you must provide. The code contains utility code to load COCO datasets as input for training.

See test cases for example usage.

• Cascade training
• Object detection

Notes

Depending on your dataset size, you'll need at least 8GB of available system memory to load image data during cascade training.

During training, scale factor must always be greater than 1.0. Otherwise, training will never end.

To convert video data:

ffmpeg -i my_video-1.mkv -t 10 \
  -crf 11 -filter:v scale=512:-1 \
  -sws_flags lanczos+accurate_rnd \
~/code/kimaris/src/test/resources/seq-00.mp4

Training image data for the picolrn binary is serialized as:

imageHeight, imgWidth, grayPixelRows, boundsCount, bounds0 ... boundsN

bounds are (r, c, s) tuples.

• r - Row - translates to an image's Y coordinate from top-left (0,0).
• c - Col - translates to an image's X coordinate from top-left (0,0).
• s - Scale - is the diameter of a circle centered at (c,r) (I think).

The reference code uses these values to represent circles of (c,r) origin, and s / 2 radius.

I have the slight suspicion that this constraints training data to have a 1:1 square aspect ratio, and generating detectors for rectangular objects like, for example, credit cards or license plates may present additional issues.

Citation

N. Markus, M. Frljak, I. S. Pandzic, J. Ahlberg and R. Forchheimer, "Object Detection with Pixel Intensity Comparisons Organized in Decision Trees", http://arxiv.org/abs/1305.4537