cmuchancel/2026-24679-image-autogluon-predictor

Water vs. No-Water Image Classifier

Purpose

This model was created for CMU 24-679 Homework 2. It performs binary image classification to predict whether a visible body of water is present in an image.

Dataset

Source dataset: ssg1/places-water-binary

Published splits were used without resplitting:

• Training: 391 images
• Validation: 5 images
• Test: 6 images

The target is:

• 1 = water
• 0 = no water

The training split contains original training photographs and augmented versions. Validation and test contain only held-out original photographs.

Input and Preprocessing

Images are 224 x 224 RGB.

The source dataset was prepared by square-cropping the photographs and resizing them to 224 x 224 pixels. AutoGluon MultiModal handles the model's expected normalization and tensor preprocessing during training and inference.

Augmentation

The training dataset includes label-preserving augmentations created from training originals only.

The augmentation methods are:

• horizontal flip
• small rotation up to approximately ±8 degrees
• brightness and contrast adjustment
• color jitter

Validation and test images are not augmented.

AutoML Search

Framework: AutoGluon MultiModal

Task: Binary image classification

Selection metric: Validation accuracy

Seed: 24679

Architectures searched:

• ResNet-18
• MobileNetV3 Small
• EfficientNet-B0

Learning rates searched:

• 0.0001
• 0.0005

This produced six architecture/learning-rate configurations.

Each trial used:

• medium_quality preset
• maximum 20 epochs
• early stopping patience of 3 validation checks
• 120-second maximum training budget per configuration

Maximum total search budget: 720 seconds.

Best Model

Best architecture: efficientnet_b0

Best learning rate: 0.0005

Best validation accuracy: 1.0000

The final model was chosen using validation accuracy only. The test set was not used to select the architecture or learning rate.

Test Results

• Accuracy: 0.5000
• Weighted F1: 0.4857

The validation split contains only 5 images and the test split contains only 6 images, so these metrics have high uncertainty. A single test error changes accuracy substantially.

Hardware / Compute

Training was performed in Google Colab using the available GPU runtime.

The AutoML search used a fixed maximum budget of 120 seconds per trial across six trials, for a maximum search budget of 720 seconds.

Limitations and Known Failure Modes

The dataset contains only 34 independently collected original photographs. The augmented images add variation but do not represent new independent scenes.

The validation and test sets are extremely small, so model selection and test metrics may vary substantially based on only one image.

The model may have difficulty when water occupies only a small portion of the image, is partially obstructed, appears under unusual lighting or weather conditions, or resembles reflective non-water surfaces.

The photographs were collected by one person using one phone, so the model may not generalize well to substantially different cameras, locations, or visual styles.

The model is intended for coursework and should not be used for safety-critical or environmental-monitoring decisions.

Ethical Considerations

The dataset is primarily composed of outdoor scenes rather than images of people. It is not intended to infer personal or sensitive attributes.

License

MIT, following the source dataset license.

Acknowledgments

The model-training workflow was adapted from the image-model course notebook provided by Professor Chris McComb for CMU 24-679, Fall 2026.

AI Usage Disclosure

ChatGPT was used to help adapt the course-provided image fine-tuning notebook into an architecture and hyperparameter search, identify the changes required for the Homework 2 rubric, and draft the initial Model Card.