students: meera & si mentor: afshin dehghan week 4: deep tracking

8
Students: Meera & Si Mentor: Afshin Dehghan WEEK 4: DEEP TRACKING

Upload: darcy-bishop

Post on 14-Jan-2016

216 views

Category:

Documents


1 download

TRANSCRIPT

Page 1: Students: Meera & Si Mentor: Afshin Dehghan WEEK 4: DEEP TRACKING

Students: Meera & SiMentor: Afshin Dehghan

WEEK 4:DEEP TRACKING

Page 2: Students: Meera & Si Mentor: Afshin Dehghan WEEK 4: DEEP TRACKING

CURRENT PROGRESS

Page 3: Students: Meera & Si Mentor: Afshin Dehghan WEEK 4: DEEP TRACKING

HANDCRAFTED FEATURES VS AUTOENCODERS

1 Online Object Tracking: A Benchmark. Yi Wu, Jongwoo Lim, and Ming-Hsuan Yang, “Online Object Tracking: A Benchmark,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013.

Downloaded 10 videos from

Online Object Tracking: A

Benchmark1, and cut them to 115

frames

Compared autoencoder

results with HOG and Color Histogram

HOG performed the worst overall

Autoencoders performed the best in all but 2

videos

Page 4: Students: Meera & Si Mentor: Afshin Dehghan WEEK 4: DEEP TRACKING

VISUALIZING THE FILTERS

Currently running code with 3 layers

Could not visualize the 2nd and 3rd layersCould visualize the 1st layer

2nd and 3rd layer representations are different

from the 1st

The current visualization function does not apply to

them

Page 5: Students: Meera & Si Mentor: Afshin Dehghan WEEK 4: DEEP TRACKING

APPLYING GAUSSIAN CONFIDENCE BASED ON MOTION VECTOR

Performance for the sequences was

improved

Change in confidence values

We observed the effect of a Gaussian motion

model

Page 6: Students: Meera & Si Mentor: Afshin Dehghan WEEK 4: DEEP TRACKING

NEXT STEPS

Page 7: Students: Meera & Si Mentor: Afshin Dehghan WEEK 4: DEEP TRACKING

Training 2

• Finish downloading 1 million images of same size

• Pre-train network with the images• Fine tune the network

Visualizing Layers

• Currently we can only visualize the 1st layer of filters

• Do more research and implement a method to visualize 2nd and 3rd layers

NEXT STEPS

2 Lamblin, Pascal and Yoshua Bengio. Important Gains from Supervised Fine-Tuning of Deep Architectures on Large Labeled Sets. NIPS 2010 Deep Learning and Unsupervised Feature Learning Workshop.

Page 8: Students: Meera & Si Mentor: Afshin Dehghan WEEK 4: DEEP TRACKING

Variations

• Patch size• Greyscale images

Further

Reading

• Fine tuning networks• Visualizing filters of higher layers• Learning motion: provide temporal data to

network so it can learn the vector

NEXT STEPS