Smoker Classification Challenge: A Computer Vision Starter Task
TL;DR
The Smoker Classification Challenge on AIOZ AI Challenges is a practical computer vision task built around binary image classification. You train a model to decide whether a person in a photo is smoking or not. With a balanced 716-image training set, 224 test images, and accuracy-based evaluation, it gives builders a focused way to practice image preprocessing, model training, validation, and submission.
Why This Challenge Matters
Smoker classification turns a simple question into a useful computer vision workflow: can a model identify smoking behavior from real-world images?
The task has practical relevance across content moderation, public health research, workplace safety, and facility monitoring. Smoking cues can be small, partially hidden, or mixed into busy scenes, so the challenge is more than a basic image exercise. It teaches the same kind of visual reasoning needed for many binary detection problems.
What You Build
You build an image classifier that predicts whether a photo shows a smoker or non-smoker.
The output is one of two labels:
- Smoker
- Non-smoker
There are no bounding boxes or segmentation masks. Each image receives one classification label, making this a clear entry point for learning how visual classification pipelines work from input image to final prediction.
Dataset Scope and Evaluation
The challenge includes 940 images in total:
- 716 training images
- 224 test images
The training set is perfectly balanced, with 358 smoker images and 358 non-smoker images. This lets participants focus on model quality rather than class-imbalance fixes.
Submissions are evaluated by accuracy, which measures the percentage of correct smoker/non-smoker predictions. Because the training classes are balanced, a model that predicts only one class starts around a 50% floor.
Workflow You Can Practice
This challenge is useful because it covers the full image classification loop in a compact setting.
You can practice:
- Loading and preprocessing image data
- Training a binary image classifier
- Using transfer learning instead of starting from scratch
- Applying safe image augmentation
- Reading training and validation curves
- Checking whether the model is learning smoking cues or background shortcuts
- Improving accuracy through controlled iteration
How to Start Efficiently
Start with a simple baseline that proves the full pipeline works before trying to optimize the model.
A practical first path:
- Review the smoker/non-smoker labels and submission format.
- Load the images and confirm the class balance.
- Run the provided baseline model to understand the workflow.
- Train a simple classifier or fine-tune a pretrained vision model.
- Submit early to establish a score.
- Improve with augmentation, validation checks, and error analysis.
Start Building
The Smoker Classification Challenge is a focused way to practice computer vision with real image data. The task is simple to understand, the dataset is small enough for fast iteration, and the balanced setup makes accuracy easy to interpret.
Join the challenge, submit a baseline, and improve your model one experiment at a time.
FAQ
Q1: What kind of machine learning task is the Smoker Classification Challenge?
It is a binary image classification task to classify each image as smoker or non-smoker.
Q2: How many images are included?
The challenge includes 940 images in total: 716 for training and 224 for testing.
Q3: Is the training set balanced?
Yes. The training set includes 358 smoker images and 358 non-smoker images.