What’s in common? cars faces pedestrians Coincidence?
Issue #1: Tougher objects Non-rectangular masks?
People Detection in the Middle Ages Nuns in Procession. French ms. ca. 1300.
Issue #2: Occlusions
e. g. PASCAL labels
The Right Spatial Support Bounding box 50 window
The Right Spatial Support Bounding box 50 window Two issues: • Which Features? • Over what support? Conjecture: getting better spatial support would allow for simpler features
Sliding Windows as Oversergmentation • Describes an image as an overlapping set of segments: • Segments are typically: • Rectangular • Axis parallel • Set of scales Why not search over shape as well? • Too much work…
Mid-level cues (segmentation) Image 12 segmentations (96 overlapping segments) -Tu et al, DD-MCMC 2001, Hoiem et al, Geometric Context, 2005; Russell et al, Object Discovery 2006
Proving my conjecture… Conjecture: getting better spatial support would allow for simpler features 1. Do multiple segmentations give you better spatial support? 2. Does better support actually help recognition? 3. Very fresh unpublished work with Tomasz
How well do segments cover objects?
Object Covergate (NCut on MSRC)
Does better spatial support help? Ground Truth Regions vs. Bounding Boxes