A 10-class image classifier that identifies which video game is being played from a gameplay screenshot. Built on a fine-tuned google/efficientnet-b0 backbone, trained at a custom, aspect-ratio-preserving 180×320 input resolution (instead of the standard 224×224 square crop) on the Bingsu/Gameplay_Images dataset.

This model was built as part of a university course project (AI Lab, SE334) — "Automated Video Game Recognition and Hashtag Suggestion for Live Streaming Platforms Using Image Classification" — and powers the GameSense demo app.

Authors: Afrim Hossen Khan

Model Details

Base model: google/efficientnet-b0
Task: Multi-class image classification (10 classes)
License: MIT
Architecture: EfficientNet-B0 backbone (ImageNet-pretrained), fine-tuned end-to-end with the final classifier layer replaced for 10 output classes. Trained at a custom 180×320 input resolution — half of the source dataset's native 640×360, preserving the true 16:9 aspect ratio — made possible without architectural changes since EfficientNet's AdaptiveAvgPool2d head is resolution-agnostic.
Fine-tuning objective: Cross-entropy loss with label smoothing (0.1), sklearn balanced class weights applied in the loss (the source dataset is already perfectly balanced at 1,000 images/class)
Training regime: Mixed-precision (AMP) training on dual CUDA T4 GPUs, AdamW optimizer with a OneCycleLR schedule, up to 25 epochs with early stopping (patience = 6, monitored on validation loss)

Classes

Among Us, Apex Legends, Fortnite, Forza Horizon, Free Fire, Genshin Impact, God of War, Minecraft, Roblox, Terraria

Intended Use

This model is intended for identifying which video game is shown in a gameplay screenshot. Example use cases:

Auto-generating hashtags/tags for gameplay clips, stream thumbnails, and social posts
Categorizing or organizing gameplay footage/screenshots by game on a content platform
A component in a larger stream metadata or content-tagging pipeline
Research and coursework on multi-class visual classification

Out of scope: This model only recognizes the 10 games listed above — any other game will be forced into one of these 10 labels rather than correctly rejected. It has been evaluated on one dataset only, and has not been validated against real-world production streaming footage, unusual camera angles, menu/loading screens, or extensive in-game cosmetic content (e.g. crossover skins) that may visually resemble a different game in the label set.

How to Use

This model is distributed in two formats — pick whichever fits your stack.

Option A: ONNX (lightweight, CPU-friendly)

Download both files and keep them in the same folder — the .onnx graph loads its weights from the .onnx.data file alongside it at runtime:

efficientnet_b0_gameplay.onnx — the ONNX graph
efficientnet_b0_gameplay.onnx.data — the external weights file

Install dependencies:

pip install onnxruntime huggingface_hub pillow numpy

Single-image prediction

python

import numpy as np
import onnxruntime as ort
from PIL import Image
from huggingface_hub import hf_hub_download

REPO_ID = "afrim3000/Game"
IMG_SIZE = (320, 180)  # PIL resize takes (width, height)
IMAGENET_MEAN = np.array([0.485, 0.456, 0.406], dtype=np.float32)
IMAGENET_STD  = np.array([0.229, 0.224, 0.225], dtype=np.float32)
CLASS_NAMES = ['Among Us', 'Apex Legends', 'Fortnite', 'Forza Horizon', 'Free Fire',
               'Genshin Impact', 'God of War', 'Minecraft', 'Roblox', 'Terraria']

# Downloads both files into the same local cache folder — required, since the
# .onnx graph references .onnx.data by relative path at load time.
onnx_path = hf_hub_download(repo_id=REPO_ID, filename="efficientnet_b0_gameplay.onnx")
hf_hub_download(repo_id=REPO_ID, filename="efficientnet_b0_gameplay.onnx.data")

session = ort.InferenceSession(onnx_path, providers=["CPUExecutionProvider"])
input_name = session.get_inputs()[0].name
output_name = session.get_outputs()[0].name

def preprocess_pil(img: Image.Image) -> np.ndarray:
    img = img.convert("RGB").resize(IMG_SIZE)
    arr = np.asarray(img, dtype=np.float32) / 255.0        # HWC, [0,1]
    arr = (arr - IMAGENET_MEAN) / IMAGENET_STD              # normalize, same stats as training
    return arr.transpose(2, 0, 1)                           # HWC -> CHW

def softmax(x: np.ndarray) -> np.ndarray:
    e = np.exp(x - x.max(axis=1, keepdims=True))
    return e / e.sum(axis=1, keepdims=True)

def predict(image_path: str):
    image = Image.open(image_path)
    x = preprocess_pil(image)[np.newaxis, ...].astype(np.float32)
    logits = session.run([output_name], {input_name: x})[0]
    probs = softmax(logits)[0]
    top_idx = int(probs.argmax())
    return CLASS_NAMES[top_idx], probs

label, probs = predict("path/to/screenshot.jpg")
print(f"Prediction: {label}")
for name, p in sorted(zip(CLASS_NAMES, probs), key=lambda t: -t[1]):
    print(f"  {name:<16} {p*100:5.1f}%")

Batch prediction

python

image_paths = ["shot1.jpg", "shot2.jpg", "shot3.jpg"]

batch = np.stack([preprocess_pil(Image.open(p)) for p in image_paths]).astype(np.float32)
logits = session.run([output_name], {input_name: batch})[0]
probs = softmax(logits)
preds = probs.argmax(axis=1)

for path, pred, p in zip(image_paths, preds, probs):
    print(f"{path}: {CLASS_NAMES[int(pred)]}  ({p[int(pred)]*100:.1f}%)")

For GPU inference, install onnxruntime-gpu instead and pass providers=["CUDAExecutionProvider", "CPUExecutionProvider"] when creating the session.

Option B: PyTorch (.pth checkpoint)

Download the checkpoint:

efficientnet_b0_gameplay_final.pth

Install dependencies:

pip install torch torchvision huggingface_hub pillow numpy

Single-image prediction

python

import torch
import torch.nn as nn
import numpy as np
from torchvision import models, transforms
from PIL import Image
from huggingface_hub import hf_hub_download

REPO_ID = "afrim3000/Game"
IMG_SIZE = (180, 320)  # (H, W) — torchvision transforms convention
CLASS_NAMES = ['Among Us', 'Apex Legends', 'Fortnite', 'Forza Horizon', 'Free Fire',
               'Genshin Impact', 'God of War', 'Minecraft', 'Roblox', 'Terraria']

ckpt_path = hf_hub_download(repo_id=REPO_ID, filename="efficientnet_b0_gameplay_final.pth")
checkpoint = torch.load(ckpt_path, map_location="cpu")

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

model = models.efficientnet_b0(weights=None)
in_features = model.classifier[1].in_features
model.classifier[1] = nn.Linear(in_features, len(CLASS_NAMES))
model.load_state_dict(checkpoint["model_state_dict"])
model.to(device).eval()

transform = transforms.Compose([
    transforms.Resize(IMG_SIZE),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

@torch.no_grad()
def predict(image_path: str):
    image = Image.open(image_path).convert("RGB")
    x = transform(image).unsqueeze(0).to(device)
    logits = model(x)
    probs = torch.softmax(logits, dim=1)[0]
    top_idx = int(probs.argmax())
    return CLASS_NAMES[top_idx], probs.cpu().numpy()

label, probs = predict("path/to/screenshot.jpg")
print(f"Prediction: {label}")
for name, p in sorted(zip(CLASS_NAMES, probs), key=lambda t: -t[1]):
    print(f"  {name:<16} {p*100:5.1f}%")

Batch prediction

python

from torch.utils.data import Dataset, DataLoader

class ImageListDataset(Dataset):
    def __init__(self, paths, transform):
        self.paths = paths
        self.transform = transform

    def __len__(self):
        return len(self.paths)

    def __getitem__(self, i):
        img = Image.open(self.paths[i]).convert("RGB")
        return self.transform(img), self.paths[i]

image_paths = ["shot1.jpg", "shot2.jpg", "shot3.jpg"]
loader = DataLoader(ImageListDataset(image_paths, transform), batch_size=8)

model.eval()
with torch.no_grad():
    for images, paths in loader:
        images = images.to(device)
        logits = model(images)
        probs = torch.softmax(logits, dim=1)
        preds = probs.argmax(dim=1)
        for path, pred, p in zip(paths, preds, probs):
            print(f"{path}: {CLASS_NAMES[int(pred)]}  ({p[int(pred)]*100:.1f}%)")

Training Data

The model was fine-tuned on the Bingsu/Gameplay_Images dataset — 10,000 gameplay screenshots (1,000 per class) at native 640×360 resolution, PNG format.

Labels: 10 classes (see Classes above)
Splits: Stratified 70 / 15 / 15 train / validation / test (the source dataset ships a single train split only; the split above was carved out manually, preserving per-class balance)
Preprocessing: Resize to 180×320 (custom, aspect-ratio-preserving resolution), ImageNet normalization (mean [0.485, 0.456, 0.406], std [0.229, 0.224, 0.225])
Training augmentation: Random horizontal flip, color jitter, random rotation (±8°), random erasing
Class balancing: The dataset is already perfectly balanced (1,000 images/class); sklearn balanced class weights are still computed and applied in the loss as a safeguard

Training Procedure

Framework: PyTorch
Hardware: Kaggle free-tier T4 x2 GPUs
Loss: Cross-entropy with label smoothing (0.1)
Mixed precision: Enabled (AMP)

Evaluation

Evaluated on the held-out test split (n = 1,500) at a decision threshold of 0.5.

Classification Report

Class	Precision	Recall	F1-score	Support
Among Us	1.0000	1.0000	1.0000	150
Apex Legends	1.0000	0.9933	0.9967	150
Fortnite	1.0000	1.0000	1.0000	150
Forza Horizon	1.0000	1.0000	1.0000	150
Free Fire	1.0000	1.0000	1.0000	150
Genshin Impact	0.9934	1.0000	0.9967	150
God of War	1.0000	1.0000	1.0000	150
Minecraft	1.0000	1.0000	1.0000	150
Roblox	1.0000	1.0000	1.0000	150
Terraria	1.0000	1.0000	1.0000	150
accuracy			0.9993	1,500
macro avg	0.9993	0.9993	0.9993	1,500
weighted avg	0.9993	0.9993	0.9993	1,500

Test ROC-AUC: 1.0000 (macro average; per-class AUC is also 1.0000 across all 10 classes)

Confusion Matrix

ROC Curve

Limitations

Performance is reported on a single dataset; generalization to other capture sources, image qualities, camera angles, or game versions/UI updates is not guaranteed.
The classifier is closed-set — it will always assign one of the 10 trained classes, even to games or content it has never seen, rather than rejecting out-of-distribution input.
Confidence can be lower on visually ambiguous content, such as games with extensive cosmetic/skin systems whose art style can resemble another class in the label set.
The model has not been evaluated as a standalone production guardrail; low-confidence predictions should be handled with a confidence threshold or human review rather than trusted outright.

Citation

If you use this model, please cite this repository and reference this course project:

@misc{game-detection-classifier,
  title  = {Automated Video Game Recognition and Hashtag Suggestion for Live Streaming Platforms
            Using Image Classification},
  author = {Afrim Hossen Khan},
  year   = {2026},
  note   = {Course project, AI Lab (SE334), Daffodil International University}
}