Hello Greg,
Here are the answers to your questions:
1) .NET wrapping: There is only .NET 2.0 support now - will it be extended to next versions? When?
No, we do not have such plans in near future. But as I know there shouldn't be any issues using .NET 2.0 with .NET 4.0.
2) Is there some benchmarking of the recognition\learning processes performance - say results of speed of recognition for models containing many submodels? with both High and Low Speeds? on different processors? does the kind of video card plays?
The only test of speed was performed gave the following results (tested on i7-920 processor):
Learning of 1 image takes 0.1261 sec.
There were used 36 images for 1 object - 4.5409 sec.
Matching speed 395 images/second
Test was performed using fast speed. Low speed is not predictable - it could be 2 or 3 times slower.
We have never performed tests with video cards.
Is type of image loaded to model playing role in performance rate?
If the image resolution is different the speed could decrease till square progression.
The bigger number of images leads to linear speed decrease.
3) What are exactly image formats the SentiSight supports? I noticed that TIFF format bitmaps fails, for example. Is there functionality to check the bitmap before it loaded - if its format fits? Will it be extended to all .NET System.Drawing.Imaging.ImageFormats? When?
SentiSight supports BMP, GIF, Nist IHead, JPEG, Lossless JPEG, JPEG 2000, PNG, TIFF and WSQ. You can check it on documentation provided with SDK (section "4.4.2 Image Format").
What regards TIFF image, could you please send it to me to check it?
To check image you can try to load from image and if you get NotSupportedException then this image is not supported.
4) The SEShape has some hidden functionality inside not described very well in the SDK manual - for example property IsValid - does it check Convexity of the shape? What does the Center property?
We know that documentation is our weak part but we are working on that.
SEShapeIsValid returns if shape is self intersected.
Center property returns the geometric center of the object.
5) If I put treshold of the engine recognition to 0 - does it mean that all recognition models will work? With what similarity score?
If you set threshold to 0 it will return all objects that are a little bit similar to the object you want (it should pass internal thresholds) and then you will have to check results by yourself (or you can use the one with the biggest score).
In short can I build some model to get list of all models recognized sorted by similarity score?
Yes, you can.