Gathering a dataset of patrons for a CV system at a shooting range in the USA

AI & Machine Learning, Data Parsing — incorrectly specified categories?

Project translated automatically. Log in or register, to view the original

The task is very interesting and original, it will appeal to the guys :)

I need to collect a dataset for training a computer vision model that will recognize 9mm Parabellum cartridges:

This is the most common cartridge in US shooting ranges

The 9*19mm Parabellum (also known as 9mm Luger, 9mm

NATO or simply 9mm)

If you think that such a dataset has value for military needs: you are deeply mistaken, because such cartridges are not counted and are supplied in buckets.

I need photos of these cartridges in different conditions, on different backgrounds, surfaces, under different lighting, etc., for training a computer vision model. (Long-term goal: to locate the cartridge, pick it up with a manipulator, place it in a box / remove it from the box, and throw it into a basket)

The ideal final result is a large number of unique photos with annotations in OBB or 6-DOP format
But overall, we can agree on any cooperation formats if you have the ability to do AABB annotations in a semi-automatic mode, that is also suitable.

Important point: cartridges can be brass or steel, the color can vary from light (new) to dark, and depending on lighting, they may reflect. Therefore, the task of accurately recognizing them is not simple.

You can scrape photos from the internet. But this will be low quality, so it can be discussed, but the value will be significantly lower. What is needed is a high-quality solution for creating and collecting the dataset.
and
Low priority:
Other cartridges are also generally interesting, but the priority is much lower.

Proposals 1 Discussions 4

Serhii Terokhin 16 May 2025

>увидеть где находится картридж и поднять его манипулятором и поставить в коробочку/ вытащить из коробочки и бросить в корзину

Если этот процесс не универсален (не на продажу манипуляторов), то, возможно, относительно простым решением было бы снять такой процесс на видео вручную, под разным освещением, ракурсами и положениями гильз. С видео создать датасет (кадры), разметить и обучить ту же Yolo или другие вижн модели.