WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

14
WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJ AND RO TOR ROELL A & AMIR R. ZAMI R

Transcript of WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

Page 1: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

WEEK 7:

WEB-ASSISTED OBJECT

DETECTION

A L E J A N D R O TO R R O E L L A &

AM

I R R . ZA M

I R

Page 2: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

GEOMETRY METHOD RESULTS

• Made threshold extremely low for each class• Needed to make sure that the true positives were

detected• Sifted through the many bounding boxes by using the GIS

arrangement to get rid of obvious spatially incorrect detections. Ex: A trash can was detected on the left side of the image when there aren’t any according to the GIS data.

• Got rid of bounding boxes that were smaller than and larger than a certain percentage of the image.

• Had to manually set GIS data for each image, crude methods of obtaining them manually didn’t give good results.

Page 3: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

GEOMETRY METHOD RESULTS:IMAGE 1• GIS fusion with three classes.• Classes were fire hydrants, street lights, and traffic lights.• Got 1/4 street lights

• The other street lights didn’t come up in the detectors at all. Too small in the image, or were too occluded

• Got 2/3 traffic signals• Got 0/1 fire hydrants

• The detector didn’t get the true positive in the image

Page 4: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

Before GIS fusion

Page 5: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

After GIS fusion

Page 6: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

GEOMETRY METHOD RESULTS:IMAGE 2 • GIS fusion with two classes.• Classes were trash cans and street lights.• Got 2/3 street lights

• The last street light didn’t come up in the detectors at all. Too small in the image and/or threshold not low enough

• Got ½ trash cans• The other trash can didn’t come up in the detectors at all.

Too small in the image and/or threshold not low enough

Page 7: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

Before GIS fusion

Page 8: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

After GIS fusion

Page 9: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

GEOMETRY METHOD RESULTS:IMAGE 3 • GIS fusion with two classes.• Classes were traffic signals and street lights.• Got 2/3 street lights• Got 3/6 traffic signals

• The other traffic signals were so close together than the detector was drown off.

Page 10: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

Before GIS fusion

Page 11: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

After GIS fusion

Page 12: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

GEOMETRY METHOD:CONCLUSIONS• Sifting the bounding boxes using the GIS data resulted in

better results compared to sifting them by their size relative to the image.

• Lowering threshold helped a lot too.• Detectors aren’t getting all the true positives which makes the

GIS fusion fail• Can’t improve what’s already failed.

• Need to implement some sort of vertical constraint to further improve results

• Had to manually set GIS data to get the best results.

Page 13: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

GOALS FOR NEXT WEEK

• Look into more automatic methods for obtaining• Field of vision• Range of visibility• Orientation of the camera

• Look into some sort of y-direction constraint• Look into using early fusion as well as late fusion

Page 14: WEEK 7: WEB-ASSISTED OBJECT DETECTION ALEJANDRO TORROELLA & AMIR R. ZAMIR.

THANK YOUFIN.