How to simulate a valid Mstar datasets？

[skimtea]

@RasmusHaapaniemi Hi RasmusHaapaniemi,i really appreciate you making the RaySAR_Python code open source. During this time, I try to generate simulated MSTAR dataset for deep learning task after referring your master thesis and several papers about RaySar. However, I encountered some difficulties, which caused the quality of my simulated images to be unsatisfactory. There may be some problems in the simulation process, but I don't know what the specific problem is, so I will show the whole process below, I wish you could help me find the problem.

1. Obtain model (T72) and generate POV file (RT) based on Accutrans 3D 3D model image

Open *.obj file in Accutrans 3d software

2. Use Povray Editer to define target and ground properties (3D position, material parameter) Camera&light_source parameter settings

Ground parameters settings

T72-tank parameter settings I tried to lower the reflection coefficient of the ground to improve the contrast between the ground and the T72 in the simulated Sar image. However, my simulation results were not in line with expectations, and there was a gap with the real MSTAR dataset

3. Obtain the contribution. txt and use Raysar Python for post-processing to generate SAR simulation images Render image

Contribution.txt

4. final simulation result （background image） （foreground image）
   （Foreground & Background Composite Image）

The noise characteristics of background ground are difficult to simulate. In your paper, you proposed the use of POvray Granite command, which can effectively simulate ground & speckle; So, How do I set target-specific parameters? (T72& Background Features) Thank you very much!!

[heqishan]

I have met the same problem with you on the background simulation. I haven't found corresponding solutions, but I think some parameters of the amplitude and scaling need to be added when using Granite pigment to adjust generated background image, which is also mentioned in the thesis. Maybe I guess this is the official way to do it.

I have tried another way to simulated images with more realistic background. That is to make a randomly fluctuating ground plane model in 3D software. Then I transformed it to .pov and merge it as the ground into the tank target .pov file. The simulation images with origin plane and generated plane are shown as following

This way can obtain some improvement but the different between my result and RasmusHaapaniemi's result still exists. If you have found better solutions, we can make a further discussion :)