Model Description
This model detects dogs in images and categorizes them into small, medium, and large based on the average weight of an adult of that breed. The weight classes generally follow: less than 30lbs for small dogs, between 30 lbs and 50lbs for medium dogs, and greater than 50lbs for large dogs. Training was done by fine-tuning the YOLOv11 model. Due to the large physical differences between dog breeds, this model is intended to be used to determine counts of each type to better meet the needs of the group in the area.
Training Data
This model is trained using the following Roboflow dataset: Link.
The Roboflow page is using a subset of the Stanford Dogs Dataset. The subset consists of images of 60 breeds across 9884 images, about half the breed and image count of the original dataset.
Annotations included manually sorting each of the 60 breeds into a category based on weight (as detailed above). Additionally, some classes were deleted due to the large weight ranges of the breed. For example, Xoloitzcuintles are usually broken into three sub-breeds with different sizes, but they are labeled in the dataset under one category.
Class Breakdown
| Metric | Small | Medium | Large |
|---|---|---|---|
| Percent | 39% | 37% | 24% |
| Count | 4,058 | 3,860 | 2,524 |
Training Split
| Metric | Train | Validate | Test |
|---|---|---|---|
| Percent | 87% | 8% | 4% |
| Count | 20,731 | 1,975 | 988 |
Training Prcedure
- Trained on Google Collab using A100
- Limited to 200 epochs and 100 patience
- Ran for 73 epochs, best at 63
- 4.9 hours of training on ~24k images
- 10k base
- 14k augmented on exposure and blur
Evaluation Resuls
Comprehensive Metrics
| Metric | Goal | Result |
|---|---|---|
| F1 Score | .75 < | .84 |
| Recall | .75 < | .96 |
| Precision | .80 < | 1.0 |
Detailed Per-Class Breakdown
Less than a .04 difference between classes for each metric.
Examples Of Classes
Visualizations
Performance Analysis
This model had high metrics across each of the classes, meeting the success threshold in precision, recall, and F1 score. The confusion matrix shows some slight overguessing, as each of the classes had a 25% to 40% rate of being predicted when that area was actual background. The model also predicted small dogs as large dogs 10% of the time, which was right at the limit set before training. That being said, the matrix still has high values of 80%-85% along the true positive diagonal. The 100% precision peak at 100% confidence does raise some red flags. This is addressed in the Known failure cases section.