Goat Detection Models — YOLOv11 / YOLOv12 (Custom Trained)
Custom-trained YOLO models for real-time goat detection on farm CCTV footage. Built for the goat-detection-counting-pipeline project — a production-grade livestock counting system developed for Deonar abattoir, Mumbai (world's largest animal market).
Model Variants
| File | Architecture | Image Size | Batch | LR | Optimizer | Best Epoch | mAP50 | mAP50-95 |
|---|---|---|---|---|---|---|---|---|
goat_yolo11n_img1024_bs16_lr0.0033_sgd_best.pt |
YOLOv11-nano | 1024 | 16 | 0.0033 | SGD | 189 / 197 | 98.99% | 78.05% |
goat_yolo11s_img1024_bs12_lr0.0025_sgd_best.pt |
YOLOv11-small | 1024 | 12 | 0.0025 | SGD | 75 / 115 | 99.05% | 79.00% |
goat_yolo12n_img1024_bs12_lr0.0075_sgd_best.pt |
YOLOv12-nano | 1024 | 12 | 0.0075 | SGD | 162 / 202 | 99.04% | 78.12% |
goat_yolo12s_img1024_bs12_lr0.0025_sgd_best.pt |
YOLOv12-small | 1024 | 12 | 0.0025 | SGD | 73 / 110 | 99.09% | 79.43% |
Recommended for production:
goat_yolo11n_img1024_bs16_lr0.0033_sgd_best.ptBest real-world performance in the live pipeline (tested on-site at Deonar). Fastest inference, lowest VRAM usage, nearly identical mAP to larger models.
Dataset
- ~20,000 annotated images — collected and annotated manually by the project team over 2+ months
- Annotation tool: CVAT (self-hosted on a $5/month VPS with Tailscale private network)
- Platform: Trained on Kaggle (P100 GPU)
- Scene diversity: top view, side view, inclined chute, vertical chute, horizontal chute, night conditions, dust, morning light, mixed animals, empty frames, garbage-heavy backgrounds
- Augmentation: mosaic (0.2), mixup (0.1), HSV shift, horizontal flip (0.4), scale (0.10), erasing (0.2), rotation (±5°)
- Classes: 1 (
goat) - Split: standard train/val split
Training Configuration
All models trained using Ultralytics with the following shared config:
from ultralytics import YOLO
model = YOLO("yolo11n.pt") # swap for yolo11s / yolo12n / yolo12s
model.train(
data="data.yaml",
imgsz=1024,
epochs=300,
batch=16, # 12 for s/n variants except yolo11n
optimizer="SGD",
lr0=0.0033, # varies per model (see table above)
momentum=0.937,
weight_decay=0.0005,
patience=20, # early stopping
single_cls=True,
pretrained=True,
amp=True, # mixed precision
cache="disk",
workers=8,
# Augmentation
hsv_h=0.0, hsv_s=0.15, hsv_v=0.25,
degrees=5.0, translate=0.05, scale=0.10,
fliplr=0.4, flipud=0.0,
mosaic=0.2, mixup=0.1, erasing=0.2,
)
Per-Model Results
All metrics are on the validation set at the best checkpoint epoch.
YOLOv11-nano (goat_yolo11n_img1024_bs16_lr0.0033_sgd_best.pt)
Training: 197 epochs total, early stopped at best epoch 189 — total training time ~4.4 hours
| Metric | Value |
|---|---|
| Precision | 96.32% |
| Recall | 96.61% |
| mAP@50 | 98.99% |
| mAP@50-95 | 78.05% |
| Val Box Loss | 0.686 |
| Val Cls Loss | 0.726 |
| Val DFL Loss | 0.859 |
Convergence progression:
| Epoch | Precision | Recall | mAP@50 | mAP@50-95 |
|---|---|---|---|---|
| 1 | 90.64% | 86.68% | 94.29% | 60.76% |
| 10 | 93.91% | 92.45% | 97.65% | 70.30% |
| 50 | 95.98% | 96.00% | 98.89% | 76.87% |
| 189 (best) | 96.32% | 96.61% | 98.99% | 78.05% |
YOLOv11-small (goat_yolo11s_img1024_bs12_lr0.0025_sgd_best.pt)
Training: 115 epochs total, early stopped at best epoch 75 — total training time ~5.1 hours
| Metric | Value |
|---|---|
| Precision | 96.69% |
| Recall | 96.75% |
| mAP@50 | 99.05% |
| mAP@50-95 | 79.00% |
| Val Box Loss | 0.660 |
| Val Cls Loss | 0.673 |
| Val DFL Loss | 0.836 |
Convergence progression:
| Epoch | Precision | Recall | mAP@50 | mAP@50-95 |
|---|---|---|---|---|
| 1 | 92.44% | 89.33% | 96.17% | 63.77% |
| 10 | 95.01% | 94.31% | 98.44% | 73.31% |
| 50 | 96.33% | 96.60% | 99.02% | 78.73% |
| 75 (best) | 96.69% | 96.75% | 99.05% | 79.00% |
YOLOv12-nano (goat_yolo12n_img1024_bs12_lr0.0075_sgd_best.pt)
Training: 202 epochs total, early stopped at best epoch 162 — total training time ~9.2 hours
| Metric | Value |
|---|---|
| Precision | 96.61% |
| Recall | 96.60% |
| mAP@50 | 99.04% |
| mAP@50-95 | 78.12% |
| Val Box Loss | 0.681 |
| Val Cls Loss | 0.690 |
| Val DFL Loss | 0.851 |
Convergence progression:
| Epoch | Precision | Recall | mAP@50 | mAP@50-95 |
|---|---|---|---|---|
| 1 | 90.64% | 86.49% | 94.56% | 57.15% |
| 10 | 93.26% | 92.27% | 97.23% | 68.70% |
| 50 | 95.87% | 96.10% | 98.93% | 76.98% |
| 162 (best) | 96.61% | 96.60% | 99.04% | 78.12% |
YOLOv12-small (goat_yolo12s_img1024_bs12_lr0.0025_sgd_best.pt)
Training: 110 epochs total, early stopped at best epoch 73 — total training time ~7.8 hours
| Metric | Value |
|---|---|
| Precision | 96.50% |
| Recall | 96.97% |
| mAP@50 | 99.09% |
| mAP@50-95 | 79.43% |
| Val Box Loss | 0.658 |
| Val Cls Loss | 0.854 |
| Val DFL Loss | 0.882 |
Convergence progression:
| Epoch | Precision | Recall | mAP@50 | mAP@50-95 |
|---|---|---|---|---|
| 1 | 93.13% | 91.39% | 97.06% | 66.74% |
| 10 | 95.09% | 94.06% | 98.31% | 72.44% |
| 50 | 96.35% | 96.77% | 99.07% | 79.03% |
| 73 (best) | 96.50% | 96.97% | 99.09% | 79.43% |
Model Comparison Summary
| Model | Params | mAP@50 | mAP@50-95 | Best Epoch | Train Time | Recommended For |
|---|---|---|---|---|---|---|
| YOLOv11-nano | ~2.6M | 98.99% | 78.05% | 189/197 | ~4.4h | ✅ Production (live CCTV) |
| YOLOv11-small | ~9.4M | 99.05% | 79.00% | 75/115 | ~5.1h | High accuracy offline |
| YOLOv12-nano | ~2.6M | 99.04% | 78.12% | 162/202 | ~9.2h | Comparison/research |
| YOLOv12-small | ~9.4M | 99.09% | 79.43% | 73/110 | ~7.8h | Best accuracy overall |
Key insight: All four models converge to nearly identical mAP@50 (~99%), demonstrating the quality of the 20k dataset. The nano models achieve 98.99–99.04% — within 0.1% of the small models — making them the clear choice for real-time edge deployment.
Live Pipeline Performance
Tested on-site at Deonar abattoir using GTX 1650 (4GB VRAM), 8GB RAM:
| Metric | Value |
|---|---|
| Inference latency p50 | |
| Inference latency p90 | ~156 ms |
| Inference latency p95 | ~178 ms |
| On-site miss rate | < 5% on 180-goat sessions |
| Stream type | Live RTSP from farm CCTV |
How to Use
With the counting pipeline (recommended)
git clone https://github.com/ubada11/goat-detection-counting-pipeline
cd goat-detection-counting-pipeline
# Download the recommended model
pip install huggingface_hub
python -c "
from huggingface_hub import hf_hub_download
hf_hub_download(
repo_id='ubada11/goat-detection-yolov11',
filename='goat_yolo11n_img1024_bs16_lr0.0033_sgd_best.pt',
local_dir='models/'
)
"
# Update configs/config.yaml:
# paths:
# weights: goat_yolo11n_img1024_bs16_lr0.0033_sgd_best.pt
python main.py
Standalone inference
from ultralytics import YOLO
from huggingface_hub import hf_hub_download
path = hf_hub_download(
repo_id="ubada11/goat-detection-yolov11",
filename="goat_yolo11n_img1024_bs16_lr0.0033_sgd_best.pt",
local_dir="models/"
)
model = YOLO(path)
results = model.predict("your_video.mp4", conf=0.25, iou=0.45)
Project Context
This model was developed as part of an R&D project for MI Tradings & General Suppliers (a BMC contractor) to automate livestock counting at Deonar abattoir, Mumbai — replacing manual head-counting with a real-time AI pipeline.
The project involved 12 months of work: on-site visits to study camera angles and chute geometry, 2+ months of data collection across all lighting and weather conditions, manual annotation of 20,000 images using CVAT, iterative training on Kaggle, and live on-site testing with real CCTV feeds.
The full pipeline includes dual-line crossing detection, ByteTrack multi-object tracking, a slot management REST API for per-vendor counting sessions, and a WebRTC live preview server.
Team: Ubada Ghawte (lead engineer & sole coder), Adil, Raafe Guide: Prof. Farhan (Rizvi College of Engineering, Mumbai) Duration: 2024–2025
License
MIT — free to use, modify, and distribute.