Object Detection Progress

The object detection of our "I" pattern card has been quite successful. Below are some pictures and details about the training process:

15 Sept 07
1) 115 pictures were collected as positive samples. For each picture, an average of 4-5 were mark out for sampling using haartraining's objectmarker.exe. In total, 468 samples were created. 1000 negatives samples (with arbitary background) were used.

Some variables such as i) background color and lighting ii) design and color of what user is wearing iii) angle of the pattern card iv) no. of pattern card v) posture vi) distance from camera etc (sit or stand) were all taken into consideration.

Samples positive pictures for training:


Different color pattern cards used:



2) Training of data were conducted twice. First was conducted on 15 Sept, results of test is not accurate, but managed to detect the "I" pattern. In this training session, only 30 negatives image were used. In the second training conduct on 16 Sept, 1000 negatives were used instead. The training took around 13 hours to complete (16 Stages of training)! Results are quite satisfatory with minor problem (i.e detect extra "I")

Results of 1st training:


Results of second training :


A few test sample detect "extra" stuff:

Some possible solutions to address this: i) training more negatives ii) Let user decide where the "I" is exactly (See Virtual Makeover where user locate exactly where the location of their eye is)

3) Also edited the testing application (performance.cpp) to extract the X, Y-Coordinates, the width and height of the detected I enclosing box. i.e we are able to obtain ratio of real life object.