Felzenszwalb Girshick Object Detection with Part Based Model

<p>Oct 07, 2014 07:01 PM &nbsp;v clear paper &nbsp;from &nbsp;Univ of Chicago. CV group.</p>

<p>

problem: &nbsp;detect and localize &nbsp;generic objects: &nbsp;people, cars etc.</p>

<p>

variation: intra class (difft cars) and also pose, illumination, non-rigid deforms, clothes on people,

colors, etc.</p>

<p>

want mixes of multiscale deformable part models.</p>

<p>

only requires boxes for the objects.</p>

<p>

system is efficient and accurate SOTA on Pascal VOC)</p>

<p>

deformable config: &nbsp;spring-like connex btwn pairs of parts.</p>

<p>

on difficult datasets, &nbsp;simpler models outperform DPM &nbsp;eg with just &nbsp;bag of features.

we address the performance gap.</p>

<p>

bikes: &nbsp;mtn bikes vs 19th cent bikes with one big wheel.</p>

<p>

want visual grammar; &nbsp;use variable hierarch strucs.</p>

<p>

difft models care share reusable parts...</p>

<p>

rich models have diffic in training; &nbsp;can easily tains simple model susing SVM eg.

rich models use latent info.</p>

<p>

part locations are NOT labeled... so they are latent (hidden vars).</p>

<p>

Dalal-Triggs won in 2006 on PASCAL db. &nbsp;used single filter HOG.

filter applied at all positions and scales of an image.</p>

<p>

input to detecotr: &nbsp;position and a scale in an image.</p>

<p>

&nbsp;</p>

<p>

&nbsp;</p>

<p>

&nbsp;</p>

<p>

&nbsp;</p>

<p>

&nbsp;</p>

<p>

&nbsp;</p>