Computer Vision :

                                                                                                                         

Here is the way to go:  Put some healthy flower, leaf, seed, or fruit on a clean white paper (for color balancing) and take photos and send the images to flower search engine along with the following additional information to reduce the complexities of the search problem.

Many well known techniques such as contour matching, invariant shape descriptors, colors an texture analysis will be required  to solve this particular problem. In case,some ambiguities are found, a short list of probable candidates is given to the user and some more questions are posed by the search engine.

We humans have great cognitive capabilities and with certainty can detect whether a given image is a painting, cartoon sketch or a photograph. We can say with great confidence whether "The picture is real or not". But what is this "Real" and what goes in our brains to classify them with aplomb?  Can we mathematically express or quantify it ? Is the  "real" thing have larger Shannon entropy or some statistical correlations that make us classifying the images instantly or do we calculate and assemble something in higher dimensions to get this answer ? From computer vision perspective, the curiosity that I have is : Can computer classify the images with high probability (if not with certainty as human beings ) and if yes, can the algorithmic efficiency compete with human perceptual skills ?

                                                                                     
Patient specific numerical simulation of hemodynamics  requires close approximation of blood vessels instead of CAD generated geometrical models. Manual segmentation and 3D surface reconstruction of medical images take enormous amount of time and therefore, automatic or semi-automatic procedure are often used in this process. There are some inherent problems with the vessel modeling with the images (1) many images have poor contrast ratio and (2) 2D image are cluttered with numerous overlapping veins. Fortunately, there are some heuristics that may make this 3D reconstruction easy (1)  vessels are smooth with no sharp corners (2) cross sectional shapes are generally tubular (3) multifurcation are rare in biological systems. Since medical imaging is commercially competitive field, and therefore I am sure that this problem must have been studied by various research group, but what I am looking for are the answers (1) How reliable are the automatic methods (if they exists) ? (2) What tools are required for semi-automatic process to reduce the 3D surface reconstruction time from days to hours ? (3) How the generated surfaces are verified and certified before neurosurgeons use them  in practice?

                                                                                                     
Polygonal representation is often required for interactive geometry exploration. Large 3D scanner are now commercially available that can provide very high resolution polygonal models, but they are both expensive and non-portable. Now a days digital cameras or camcorders are quite inexpensive and to achieve high resolution representation  from these gadgets a large number of images from different views may be required. Perhaps the most important things is to capture salient features (such as ridges, valleys etc) from some suitable direction so that 3D surface reconstruction algorithms capture them to produce perceptually acceptable model, also it is desirable that the entire process is done in some reasonable time and with limited resources. Interaction with the user for missing parts or to remove topological ambiguities may also be essential. The unanswerable questions are (1)  How many images are required (2) Can we really generate a polygonal model for which even a child can say:  Yes, it  looks very similar to the one I saw in the temple ?

                                                              

Recently I was tracking the path of  rivers Ganges and Brahmaputra with the Google earth manually. Although it took great patience and time, I was able to complete it and verified the path with other geographic information available. I was also able to find out that Ganges starts from somewhere Gangotri in Himalayas and Brahmaputra somewhere in Tibet plateau.  It took hours to identify the main path of the river which has vast number of small and large tributaries. The tracking was more time consuming than identifying road systems because of meandering nature of rives. I think that perhaps some automation is possible to identify the rivers paths from the Google or aerial images with some ideas taken from automatic road detection systems.  Such automatic process can be extremely useful for GIS applications, urban planning, and flood control.

                                                              

Problem: Very often 3D polygonal models are modified by some affine transformations and for reducing the complexity (mesh simplification) and sometimes by clipping or implanting something on the object, Therefore identifying the similarity between two objects both locally and globally is important.  Establishing correspondence between salient feature points, lines and contour is essential to solve the problem.  For partial similarity detection, consistent segmentation of the model is performed and correspondence in segment pairs is established.  Invariant such as surface ridges and valleys is used for both data reduction and identifying salient features.

Application: There are large number of precious statues and artifacts that are smuggled from India every year and there are few mechanism available for the law enforcement agencies to find out whether the object is part of the national treasure. Suppose there is national database of  such objects and if the officials are able to take few photograph of the objects and query for the similarities from the national antique/historical database, then it could help in combating the menace of illegal smuggling.

Project -I :  High Dynamic Range Imaging:

Project-II :  Image Stitching:

Project-III :  Photometric Stereo 

Project IV: 

Class Presentation (PDF File)

Class Term Paper (PDF)

MultiView 3D Reconstruction: Exploiting Massive Parallelism