Class 6.870 "Object Recognition and Scene Understanding" by Antonio Torralba
Evaluation of "A Parametric Texture Model Based on Joint Statistics of Complex Wavelet Coefficients":
Forest detection in high-resolution aerial images
R. Gokberk Cinbis
Motivation for the
evaluation methodology:
Due to recent advances in remote sensing systems and higher user expectations, importance of high resolution aerial image analysis has increased. However, it is also less-studied. Texture analysis is a central problem in high resolution image analysis, especially for research problems regarding tree/forest detection.
Forest texture is one of the most challenging texture patters since it varies a lot depeding on factors such as;
- Physical properties of the ground
- Illumination conditions
- Tree species
Here, I use a simple idea regarding this problem: "If a parametric model contains enough information for texture synthesis, then it can be also useful for texture classification" as many researchers did for evaluating texture models.
Dataset:
I have collected 26 cropped (128x128) pictures from several satellite images and labelled them as positive (a forest picture) or negative. The images are the following:
Please note that these images are relatively easy as they have a relatively regular texture. In addition, tiling is a well know simplification for dealing with large images, however, it is generally problematic as it does not respect semantic region boundaries. However, these images are collected to eliminate such problems.
Experiments:
I have created texture models using the code for the paper J. Portilla and E. Simoncelli. A Parametric Texture Model Based on Joint Statistics of Complex Wavelet Coefficients. IJCV 2004. Then, I divided the images into two groups: training (first 7 images in each group) and testing (last 6 images). For each test image, I have found the Nearest Neighbor using Euclidean distance metric using all statistics extracted for texture synthesis purposes.
There are three major parameters for extracting these statistics:
In the following table, NN classification results are presented. In each image pair, the left one corresponds to the test image, and the right one corresponds to its closest neighbor in the training set.
and,
We observe that both average performance and (subjective) relevance of nearest neighbors increases dramatically when we use #scales = 2. As a natural extension to this observation, I also experimented to see whether it holds for texture synthesis too. However, there is not much difference between two cases, which shows that smaller scales may provide poor information for classification purposes, however, it might be useful for synthesis purposes (as it would be expected). Results are shown as follows:
Also I have checked the synthesis results for negative examples whether it can provide any meaningful pattern, but it did not (as expected):
As suggested in the class, I have done the NN matching using pixelwise Sum of Squared Distances (SSD) for comparison (different normalization methods didn't improve the results):
Conclusions
In these experiments, we have seen that the compact representation provided in the paper by Portilla and Simoncelli for texture analysis is useful for both texture analysis and synthesis using aerial forest images. We have also seen that some data useful for synthesis can be noisy for classification (which can be handled by a more intelligent classifier) and data poor for synthesis can be useful for classification (negative images).
Oct'08
My homepage
Due to recent advances in remote sensing systems and higher user expectations, importance of high resolution aerial image analysis has increased. However, it is also less-studied. Texture analysis is a central problem in high resolution image analysis, especially for research problems regarding tree/forest detection.
Forest texture is one of the most challenging texture patters since it varies a lot depeding on factors such as;
- Physical properties of the ground
- Illumination conditions
- Tree species
Here, I use a simple idea regarding this problem: "If a parametric model contains enough information for texture synthesis, then it can be also useful for texture classification" as many researchers did for evaluating texture models.
Dataset:
I have collected 26 cropped (128x128) pictures from several satellite images and labelled them as positive (a forest picture) or negative. The images are the following:
 |
 |
Please note that these images are relatively easy as they have a relatively regular texture. In addition, tiling is a well know simplification for dealing with large images, however, it is generally problematic as it does not respect semantic region boundaries. However, these images are collected to eliminate such problems.
Experiments:
I have created texture models using the code for the paper J. Portilla and E. Simoncelli. A Parametric Texture Model Based on Joint Statistics of Complex Wavelet Coefficients. IJCV 2004. Then, I divided the images into two groups: training (first 7 images in each group) and testing (last 6 images). For each test image, I have found the Nearest Neighbor using Euclidean distance metric using all statistics extracted for texture synthesis purposes.
There are three major parameters for extracting these statistics:
- Number of scales
- Number of orientations
- Spatial neighborhood size
In the following table, NN classification results are presented. In each image pair, the left one corresponds to the test image, and the right one corresponds to its closest neighbor in the training set.
 |
| Number of scales = 4 (AP = 50%) |
and,
 |
| Number of scales = 2 (AP = 100%) |
We observe that both average performance and (subjective) relevance of nearest neighbors increases dramatically when we use #scales = 2. As a natural extension to this observation, I also experimented to see whether it holds for texture synthesis too. However, there is not much difference between two cases, which shows that smaller scales may provide poor information for classification purposes, however, it might be useful for synthesis purposes (as it would be expected). Results are shown as follows:
| Source Image | Synthesis using #scales = 2 | Synthesis using #scales = 4 |
 |
 |
 |
 |
 |
 |
Also I have checked the synthesis results for negative examples whether it can provide any meaningful pattern, but it did not (as expected):
| Source Image | Synthesis using #scales = 2 | Synthesis using #scales = 4 |
 |
 |
 |
As suggested in the class, I have done the NN matching using pixelwise Sum of Squared Distances (SSD) for comparison (different normalization methods didn't improve the results):
 |
| Piselwise SSD - NN matching |
Conclusions
In these experiments, we have seen that the compact representation provided in the paper by Portilla and Simoncelli for texture analysis is useful for both texture analysis and synthesis using aerial forest images. We have also seen that some data useful for synthesis can be noisy for classification (which can be handled by a more intelligent classifier) and data poor for synthesis can be useful for classification (negative images).
Oct'08
My homepage