---
license: other
task_categories:
  - object-detection
tags:
  - counter-strike
  - cs2
  - game
  - yolo
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train/**
      - split: validation
        path: data/validation/**
dataset_info:
  features:
    - name: image
      dtype: image
    - name: objects
      struct:
        - name: bbox
          sequence:
            sequence: float32
        - name: categories
          sequence:
            class_label:
              names:
                '0': none
                '1': ct_body
                '2': ct_head
                '3': t_body
                '4': t_head
---

# CS2 player detection - native 640x640 centre crops

5473 labelled frames from Counter-Strike 2 with 10643 boxes for
player bodies and heads, split by team.

A YOLO26 detector trained on this data is at
[fvossel/csgo-player-detection](https://huggingface.co/fvossel/csgo-player-detection). Its training set is not quite
identical: it also contains the images from the source dataset named below, which
are not redistributed here.

## What makes this different

Every image is a **640x640 centre crop at native resolution - never a resized
screenshot**. A full frame scaled down to 640 shrinks a head to a handful of
pixels and changes its apparent size with the capture resolution. Cropping instead
keeps a head exactly the pixel size it has on screen, so a model trained here
transfers to live inference at the same crop geometry. Mixing this data with
whole-screenshot datasets reintroduces precisely that scale mismatch.

## Classes

| ID | Name | train | validation | total |
|---|---|---|---|---|
| 0 | `none` | 0 | 0 | 0 |
| 1 | `ct_body` | 2110 | 523 | 2633 |
| 2 | `ct_head` | 1836 | 470 | 2306 |
| 3 | `t_body` | 2374 | 638 | 3012 |
| 4 | `t_head` | 2098 | 594 | 2692 |

Class `none` exists to keep the IDs stable and carries no instances.

## Splits

| Split | Images | Boxes | Images without a box |
|---|---|---|---|
| train | 4357 | 8418 | 900 |
| validation | 1116 | 2225 | 215 |

The split is **block-wise, not per frame**: consecutive frames from one scene stay
together on the same side. Shuffling per frame would put near-identical images in
both splits and inflate the validation score. Images without a box are kept on
purpose as negative examples.

## Box format

`objects.bbox` is COCO-style `[x, y, width, height]` in **absolute pixels** within
the 640x640 crop, with `objects.categories` parallel to it.

## Sources

This dataset started from [Counter Strike 2 Body and Head Classification](https://www.kaggle.com/datasets/merfarukgnaydn/counter-strike-2-body-and-head-classification)
by Kaggle user `merfarukgnaydn`. Its five-class scheme with the IDs
`none / ct_body / ct_head / t_body / t_head` is kept unchanged here, so labels stay
compatible with it.

No image from it is redistributed here.

All 5473 images here were collected and labelled for this dataset, from
two kinds of material:

- Live captures from offline matches.
- Frames rendered from community demos via CS Demo Manager, at 5 fps from tick
  ranges around gunfire, from both the T and the CT perspective. X-ray was off,
  so no player is visible through a wall.

Sessions: 20260724-152753-b7fd0f, 20260724-153124-024a7c, 20260725-124520-f30d8b, 20260725-125247-becb26, 20260725-185119-fc031b, demo_dust2_ct_01, demo_dust2_t_01, demo_dust2_t_02

## Labelling

Boxes were produced by a model and then **confirmed or corrected by hand, image by
image**. Only confirmed images are in this dataset; unchecked model output is not.

## Licence and derivation

The images show Counter-Strike 2, so they contain Valve's game assets and are
derivative of them. They are published for research on object detection. Check for
yourself whether your intended use is covered before you build on this. The
annotations themselves are free to use.

The Kaggle dataset that inspired this one states no licence, so none of its images are included here.

## Acknowledgement

Thanks to `merfarukgnaydn`, whose dataset supplied the class scheme this one still
uses and was the starting point for the work.