Object and Activity Recognition in a Distributed Camera Network

15 downloads 38 Views 1MB Size Report
in a Distributed Camera Network. N. Naikal, D. Singaraju and S. S. Sastry. SWARM Lab Summer Retreat 2012. 5/18/2012. Sastry Research Group. 1 ...
Object and Activity Recognition in a Distributed Camera Network N. Naikal, D. Singaraju and S. S. Sastry SWARM Lab Summer Retreat 2012.

5/18/2012

Sastry Research Group

1

Motivation



Interesting objects/landmarks and events are commonly viewed by smart camera networks. – Examples: Security cameras, smart-phones, sports/media coverage, etc.



Large volumes of visual data available.



Goal: Leveraging multiple views of common objects and events to improve recognition.

5/18/2012

Sastry Research Group

2

Experimental Setup



Berkeley Multi-view Wireless database. – 20 landmarks at UC Berkeley. – 16 vantage points (large baseline); 5 images per location (small baseline).

• Future setup – network of 50 cameras in Cory 337. 5/18/2012

Sastry Research Group

3

Outline • Distributed object recognition. – Efficient system for recognition of landmarks in wireless camera networks. – Scheme for informative feature selection.

• Human activity recognition – Model for detection of articulated objects in single view. – Propose extension to multiple views. 5/18/2012

Sastry Research Group

4

Landmark Recognition in Camera Network Given training images of landmarks in pre-defined categories; Assign category label to query images.

• System constraints – Cameras have limited processing power. – Band-limited communication channel. – No inter-camera calibration assumed. 5/18/2012

Sastry Research Group

5

Bag-of-Words (BoW) Framework

Keypoint

Keypoint

Extraction

Extraction

Codeword Quantization

Visual Words Representation 5/18/2012

Sastry Research Group

6

System Pipeline



Visual word histograms are non-negative and sparse.



Compress using random projection (Coefficients of 𝐴 ∈



ℝ𝑑×𝐷

𝒃𝑘 = 𝐴𝒙𝑘 drawn from zero-mean Gaussian distribution).

Decompress using 𝑙1 -minimization 𝒙𝒌 = arg min 𝒙

5/18/2012

1

subject to 𝒃𝒌 = 𝐴𝒙

Sastry Research Group

7

Large Baseline Results

5/18/2012

Sastry Research Group

8

Irrelevant Feature Suppression

Keypoints returned by interest point detector

• Non-useful features generally generated.

• Method needed to eliminate irrelevant features during training. Proposed Sparse Principal Components Analysis to identify informative features 5/18/2012

Sastry Research Group

9

Sparse PCA for Identifying Informative Features Sparse PCA Σ − empirical covariance matrix, 𝑟 − desired cardinality, 𝐱 − sparse PC.

Informative features correspond to non-zero support of sparse PC

Original Features 5/18/2012

Informative Features Sastry Research Group

10

Performance on BMW Database

• Union of informative features from all categories forms support set. • Total of 400 words in support set. • Visual dictionary constructed with 1000 words.

5/18/2012

Sastry Research Group

11

Models of objects with deformable parts Candidate hypothesis

(𝑦1 , 𝑦2 , … , 𝑦10 )

𝑦𝑝 : location of the 𝑝𝑡ℎ part.

Felzenszwalb et al. PAMI ‘09, Zhu et al. CVPR ‘10, Yang et al. CVPR ’11. 5/18/2012

Sastry Research Group

12

Towards Unifying Detection and Segmentation • Can we improve the results of the discriminative object detectors by unifying detection and segmentation?

Park and Ramanan. ICCV’11

• How do we define a cost function for unified detection and segmentation? 5/18/2012

Sastry Research Group

13

An Energy Function for Unified Detection and Segmentation

• 𝑑𝑝 𝑦𝑝 ∈ *0, 1+: variable for detection/occlusion of the 𝑝𝑡ℎ part at location 𝑦𝑝 . • 𝑥𝑖 ∈ *0, 1+: segmentation of the 𝑖𝑡ℎ pixel.

+

5/18/2012

Sastry Research Group

14

An Energy Function for Unified Detection and Segmentation

• 𝑑𝑝 𝑦𝑝 ∈ *0, 1+: variable for detection/occlusion of the 𝑝𝑡ℎ part at location 𝑦𝑝 . • 𝑥𝑖 ∈ *0, 1+: segmentation of the 𝑖𝑡ℎ pixel. • 𝒮𝑝 : shape prior for the 𝑝𝑡ℎ part (learned from training data).

𝒮 ℎ𝑒𝑎𝑑

5/18/2012

Sastry Research Group

15

Results from initial tests: input to our algorithm

Given hypothesis for parts’ locations

Hypothesis for location of parts (obtained using existing detectors) 5/18/2012

Probability of each pixel being assigned to the foreground (shape priors)

Sastry Research Group

16

Results from initial tests: output from our algorithm

Given hypothesis for parts’ locations

Locations of the “detected” parts (yellow) and the “rejected” parts (red) 5/18/2012

Sastry Research Group

Segmentation 17

Results from initial tests: input to our algorithm

Hypothesis for location of parts (obtained using existing detectors) 5/18/2012

Probability of each pixel being assigned to the foreground (shape priors)

Sastry Research Group

18

Results from initial tests: output from our algorithm

Locations of the “detected” parts (yellow) and the “rejected” parts (red) 5/18/2012

Sastry Research Group

Segmentation 19

Conclusion • Presented a distributed object/landmark recognition framework for wireless smart cameras. • Demonstrated effectiveness of using Sparse PCA for informative feature selection. • Presented scheme to unify segmentation and detection of articulated models in single images.

Future Work • Using anthropometry constraints to improve human detection in multiple views. • Tracking articulated models in video stream. • Recognizing human activity from reconstructed models.

5/18/2012

Sastry Research Group

20

Suggest Documents