Background removal techniques

[Sarah5567]

Adds a detailed README comparing vegetation segmentation methods used as preprocessing before nvinfer.

Includes:

• Code examples for six methods
• Accuracy, latency, and GPU support comparison
• Visual results and analysis

Result: ExG + Otsu + Morphology recommended as the final preprocessing method (best accuracy).

closes #75

[lyuzinmaxim]

Please do not include this section.

The classes we fail on should be documented, and we don't want to exclude them to get better accuracy numbers (it's an antipattern in general!).

We performed hierarchical evaluation without excluding any classes. This ensures that we can assess performance across broader categories - for example, misclassifying a two-leaf maize stage as a four-leaf stage is much less critical than misclassifying it as a weed

[Sarah5567]

But we can’t say that our model is able to classify those classes - its accuracy on them is zero.

[lyuzinmaxim]

@Sarah5567 and we don't have to say that we're able to classify then. We just need to document this as it is

[Sarah5567]

I'm not sure I understand you - is there a specific problem with the section that starts with Post-Optimization Results, or is the overall content of Class-Specific Performance Issue incorrect?