Training on thermal images

[varungupta31]

I have thermal image (colored) of size - Height 240 and width 320.

I'm planning to train a custom detector,

• what width and height values in the config file would work best? ( I understand it has to be multiple of 32, but what would probably work best)
• I read the article discussing hue, saturation, etc. in darknet but still felt confused whether I shall edit these in my config and how?

My thermal images contain a color pallet such as this:

• Will a model trained on lower res images work well on higher res?
• If I upscale the height and width, would that distort the images and lead to poor results?

Let me know how shall I optimize my config file that increase my chances of accurate results?

Thanks.

[stephanecharette]

If your images are 240x320, why not use 256x320 as your network dimensions? Or 224x320 is just as good, and slightly lower resolution so less pixels to process (which is good unless you are trying to find something small).

As for hue, sat, etc, it depends on the range you'd typically see in your images, and what you have available to train.

As for what to do about sizing...what are you looking to find? What is the size of the object compared to the size of the image? Do your training images match the images you'll be using for inference? That is what actually matters.

[varungupta31]

unless you are trying to find something small

So the object i'm trying to detect would be around 20x20 or 25x20 (something in this range), minimum dimension of detection would be around 10x15. when the image is 240x320.

Also, I'd not say that detection is going to very easy, because there quite a few images where the object of interest gets a bit blurry or get mixed with the background. (if this helps in recommending a right height x width value)

• Is this small, or I can use 256x320 for this?

Also, would 416x416 (default) be an overkill? Besides adding complexity, does it pose other disadvantages?

As for hue, sat, etc, it depends on the range you'd typically see in your images

How do I check that so I can make a better guess at setting these params (hue, etc.) if needed at all. (range in terms of color, right? the color is quite similar to image I shared in the post)

Do your training images match the images you'll be using for inference

It does.

@stephanecharette Kindly let me know, Thank you so much for the help!