Object detection training for embedded platforms.

Acf-jacinto is a modification of the Piotr's toolbox to enable training of ACF object detection models suitable for low power embedded platforms.

The following were the main changes made in acf-jacinto compared to the Piotr’s Toolbox:

• Fast HOG computation method (that doesn't use division or tabkle lookup) and cell sum. It is not MMX accelerated yet.
• Use simple to compute YUV features instead of LUV.
• Disable feature rescale and use feature computation for each scale. (This change is not really required for a reasonably large dataset captured on road, but it helps in InriaPerson dataset, when using the above HOG computation).
• Include positive images also in bootstrap.
• Change channel order and put HOG first.
• Disable gradient normalization as it involves division, and also doesn't seem to be helping.
• Do input pre-processing: adaptive histogram equalization and smoothing.
• Change default size to 64x64 feature, 24x56 object model to be able to detect farther, smaller objects.
• Change detection threshold.
• Write out the descriptor in an easy to read text format.
• Support MP4 videos for extraction. Option to extract only annoted frames in a video. Only few frames in a video need to be annotated. But if a frame is annotated, that frame should be fully annotated for teh object of interest.
• Dataset extraction made easy. Just specify your videos and vbb files directly in the matlab script.
• Added the caltech pedestrian bench marck evaluation labeling code into the repository under the folder vbb.
• vbbLabeler (see vbb folder) is updated to be able to open and annotate MP4 videos. It can also read a list of bjects if a file called objectTypes.txt is present in the current folder.
• Other cosmetic and usability improvements.

• Open Matlab and navigate to detector folder.
• Open acfJacintoExample.m in editor
• Make changes for your dataset path, list of videos and annotations files, object type to be trained etc.
• Run the file to do train and test.

The following links will direct to the original Acf / Piotr's toolbox

https://github.com/pdollar/toolbox
https://pdollar.github.io/toolbox/