Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the...

49
Deep Geolocalization and Siamese Nets Computer Vision James Hays

Transcript of Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the...

Page 1: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Deep Geolocalization and Siamese Nets

Computer Vision

James Hays

Page 2: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Long, Shelhamer, and Darrell 2014

Page 3: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Long, Shelhamer, and Darrell 2014

Page 4: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

PlaNet - Photo Geolocation with Convolutional Neural Networks

Tobias Weyand, Ilya Kostrikov, James Philbin

ECCV 2016

Page 5: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Discretization of Globe

Page 6: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Network and Training

• Network Architecture: Inception with 97M parameters

• 26,263 “categories”

• 126 Million Web photos

• 2.5 months of training on 200 CPU cores

Page 7: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,
Page 8: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,
Page 9: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

PlaNet vs im2gps (2008, 2009)

Page 10: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Spatial support for decision

Page 11: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

PlaNet vs Humans

Page 12: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

PlaNet vs Humans

Page 13: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

PlaNet summary

• Very fast Geolocalization method. Geolocalization by categorization.

• Uses far more training data than previous work (im2gps)

• There’s definitely still room for improvement

Page 14: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Learning Deep Representations For Ground-to-Aerial Geolocalization

Tsung-Yi Lin, Yin Cui, Serge Belongie, James Hays

CVPR 2015

Page 15: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Where was the photo taken?

View From Your Window ContestJune 9, 2010 – Feb. 4, 2015

Page 16: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Ans:Milano, Italy

Page 17: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

To Geolocalize a Photo

• One can capture every corner on the earth

Page 18: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

To Geolocalize a Photo

Page 19: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,
Page 20: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,
Page 21: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

How To Match Ground-to-Aerial?

Shan et al., Accurate Geo-registration by Ground-to-Aerial Image Matching, 3DV’14Bansal et al., Ultra-wide baseline façade matching for geo-localization, ECCV workshop’12

Page 22: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Are these the same location?

Ground

Aerial

Page 23: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Are these the same location?

Ground

Aerial

Page 24: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Why Don’t You Just…

• Sparse Keypoint Matching + RANSAC

Page 25: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Cross-view Pairs

Ground

Aerial

Page 26: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Heading DirectionGPS location

Street-view car

Page 27: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Heading DirectionGPS location

Street-view car

Page 28: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Heading DirectionGPS location

Street-view car

45°

Page 29: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

7 Cities: 78k Corresponding Pairs

San Francisco San Diego Chicago Charleston

Tokyo Rome Lyon

Page 30: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Place Verification

Same OR

Different

Page 31: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Face Verification

Same OR

Different

Page 32: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Face Verification

• Chopra and Hadsell and LeCun, Learning a Similarity Metric Discriminatively, with Application to Face Verification (CVPR 2005)

• Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014)

• Schroff, Kalenichenko, Philbin, FaceNet: A Unified Embedding for Face Recognition and Clustering (CVPR 2015)

Huang et al. Labeled Faces in the Wild: A Database for Studying Face Recognition in Unconstrained Environments

Page 33: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

“Siamese” ConvNet for Ground-to-Aerial Matching

Page 34: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

“Siamese” ConvNet for Ground-to-Aerial Matching

Page 35: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

“Siamese” ConvNet for Ground-to-Aerial Matching

Page 36: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Contrastive Loss

red: similar pairsblue: dissimilar pairs

Hadsell, Chopra, Yann LeCun, Dimensionality Reduction by Learning an Invariant Mapping, CVPR06

2

2

Page 37: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Pair Distance Distribution

Margin

ImageNet-CNN Model Where-CNN Model

50k iterations

Green: positive pairsRed: negative pairs

Page 38: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Quantitative Evaluation (AP)

• Random: 5% (1:20 pos. to neg. pairs)

• HoG2x2 (BoW): 7.9%

• Places-CNN: 10.2%

• ImageNet-CNN: 11.3%

• Where-CNN (ours): 41.9%

Page 39: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Share The Same Parameters?

• For face verification A and B share parameters

• For ground-aerial image pairs, should A, B share parameters?

Page 40: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Quantitative Evaluation

Page 41: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,
Page 42: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,
Page 43: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,
Page 44: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,
Page 45: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,
Page 46: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Strongest Activations of Particular Units

Page 47: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Geolocalization

San Francisco San Diego Chicago Charleston

Tokyo Rome Lyon

Page 48: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,
Page 49: Deep Geolocalization and Siamese Nets · • Taigman, Yang, Ranzato, Wolf, DeepFace: Closing the Gap to Human-Level Performance in Face Verification (CVPR 2014) • Schroff, Kalenichenko,

Conclusions

• Localize images without corresponding ground-level images

• Create a large-scale training dataset from public data sources

• Learning feature representations for matching cross-view images