Presentation is loading. Please wait.

Presentation is loading. Please wait.

Implementing GIST on the GPU. Refrence Original Work  Aude Oliva, Antonio Torralba  Modeling the shape of the scene: a holistic representation of the.

Similar presentations


Presentation on theme: "Implementing GIST on the GPU. Refrence Original Work  Aude Oliva, Antonio Torralba  Modeling the shape of the scene: a holistic representation of the."— Presentation transcript:

1 Implementing GIST on the GPU

2 Refrence Original Work  Aude Oliva, Antonio Torralba  Modeling the shape of the scene: a holistic representation of the spatial envelope  International Journal of Computer Vision, Vol. 42(3): 145-175, 2001.  CPU Implementation  http://cvcl.mit.edu/Papers/IJCV01-Oliva- Torralba.pdf Our Work  Parallelize to work on the GPU

3 Introduction Recognition of real world scenes Spatial Envelope  A very low dimensional representation of the scene is called the Spatial Envelope.  Takes care of Naturalness Openness Roughness Expansion Ruggedness

4 Intuitive Notion Naturalness  Straight lines indicate man made structures  Crooked, rough lies indicate natural environment. Openness  Presence of horizon reflects highly open environment Roughness  Refers to the size of its major components.  Is correlated with the fractal dimension of the scene

5 Intuitive Notion contd… Degree of Expansion  Converging parallel lines allows to percept depth gradient of the space.  A flat view of a building would have a low DoE  A street with long vanishing lines would have a high DoE. Degree of Ruggedness  The deviation of the ground wrt horizon  Open environments have a flat horizontal ground level  Mountainous landscapes have rugged ground  Ruggedness produce oblique contours, hide the horizon.  Man-made environments built on flat ground, less rugged.

6 Basic Algorithm Create Gabor Filter Bank Preprocess image  Local contrast normalization  Local luminance invariance normalization for color images Descriptor  Convolve image with filters  For each filter Divide image into blocks Mean of each block is a number in feature vector

7 Parameters Involved Number of Scales  Typically = 3 Number of orientations per scale  Typically = 8 Image size  Typically 320x240 to 1024x1024

8 Ways to Parallelize On Number of scales * orientations per scale  Number of threads = 24  Less than typical image size On Image MxN Pixels  Number of threads = M*N >> 24.  High degree of parallelism.

9 Parallelization for creation of Gabor filter Pixel Level Threads calculating Gabor filter value

10 Parallelization for prefiltering image Pixel Level Fast Fourier Transforms Pixel by pixel Multiplications => Inherent Parallelism

11 RGB image components Filters Descriptor Division into blocks Calculating Descriptor

12 Graphs

13 Speed up Image SizeCPU Time ( ms )GPU Time ( ms )‏ 64 x 64250 350300 128 x 128750450300 256 x 2563140800400 512 x 512122002100550 1024 x 10246550051001200

14 Thank You !!


Download ppt "Implementing GIST on the GPU. Refrence Original Work  Aude Oliva, Antonio Torralba  Modeling the shape of the scene: a holistic representation of the."

Similar presentations


Ads by Google