week 4: web-assisted object detection alejandro torroella & amir r. zamir
TRANSCRIPT
WEEK 4:
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
PRE-TRAINED DPM MODEL: BICYCLE
Images with bicycles in the frame:
PRE-TRAINED DPM MODEL: BICYCLE
Images without bicycles in the frame:
TRAINED DPM MODEL ON PASCAL VOC2012 DATASET: BICYCLE
Images with bicycles in the frame:
TRAINED DPM MODEL ON PASCAL VOC2012 DATASET: BICYCLE
Images without bicycles in the frame:
TRAINED DPM MODEL ON IMAGE-NET DATASET: TRAFFIC LIGHTS
Images with traffic lights in the frame:
TRAINED DPM MODEL ON IMAGE-NET DATASET: TRAFFIC LIGHTS
Images without traffic lights in the frame:
TRAINED DPM MODEL ON STEFFI MORRIS’ DATASET: TRAFFIC LIGHTS
Images with traffic lights in the frame:
TRAINED DPM MODEL ON STEFFI MORRIS’ DATASET: TRAFFIC LIGHTS
Images without traffic lights in the frame:
CONCLUSIONS:
• DPM model trained on the Image-Net dataset performed better than Steffi Morris’ manually annotated dataset.• Likely due to the fact that Steffi’s dataset was much smaller
(~150 vs ~1200)• I believe that both datasets can be better annotated (include
pose) to increase performance.
• DPM model I trained on the VOC2012 dataset performed ever so slightly better than the model pre-trained on the VOC2010 dataset• Makes sense since the VOC2010 dataset is a subset of the
VOC2012 dataset
GIS DATASETS: LOS ANGELES AND D.C.
Found GIS data on fire hydrant, street lights, traffic lights and bus stops for the Los Angeles county
Found GIS data for fire hydrants, metro entrances, bus stops, and AM/FM/Cell towers for Washington D.C.
Final choice of dataset to use will depend on DPM results on metro stations, street lights and AM/FM/Cell towers, which I have doubts on how well they can be detected and the quality of the training dataset that can be found on these objects.
THANK YOUFIN.