Falcon1-2.6M-Image

A lightweight object detection model with 2.6M parameters, fine-tuned on Pascal VOC.

It is based on YOLO11n. The backbone was frozen and only the neck and head were trained for 30 epochs on Pascal VOC. It achieves 85.4% mAP50 on VOC validation, which is higher than a full YOLO11n trained on the same dataset.

File: Falcon1-2.6M.pt (5.5 MB)

Metrics

All metrics are on Pascal VOC 2007 validation set. 4952 images, 12032 instances, 20 classes, image size 640.

Metric	Score
mAP50	85.4%
mAP50-95	65.4%
mAP75	68.9%
Precision	83.1%
Recall	77.4%
F1	80.1%
Fitness	67.4%
Parameters	2.6M
GFLOPs	6.5
Inference (T4, 640px)	2.3 ms
Inference (T4, 320px)	1.1 ms
Model size	5.5 MB

Fitness is calculated as 0.1 * mAP50 + 0.9 * mAP50-95.

Per-Class Results (mAP50 / mAP50-95)

Class	mAP50	mAP50-95
aeroplane	94.0%	71.2%
bicycle	88.5%	66.4%
bird	84.1%	58.9%
boat	78.2%	50.1%
bottle	74.9%	53.9%
bus	90.9%	74.5%
car	93.6%	72.8%
cat	90.8%	70.2%
chair	68.8%	48.1%
cow	86.4%	64.3%
diningtable	76.5%	57.8%
dog	89.5%	67.6%
horse	93.5%	71.9%
motorbike	90.2%	66.7%
person	91.4%	66.1%
pottedplant	62.9%	36.8%
sheep	85.7%	62.4%
sofa	80.3%	60.9%
train	91.8%	68.4%
tvmonitor	84.6%	61.5%

Strongest classes: aeroplane, car, horse, bus, person.

Hardest classes: pottedplant, chair, bottle. Small and cluttered objects are the main weakness, which is typical for lightweight detectors.

Comparison With Other Models

All models evaluated on Pascal VOC validation, image size 640, same protocol where available.

Model	Parameters	mAP50	mAP50-95	Size	Speed (T4)
Falcon1-2.6M (this model)	2.6M	85.4%	65.4%	5.5 MB	2.3 ms
YOLO11n (full training)	2.6M	83.5%	62.5%	5.5 MB	2.3 ms
YOLOv10n	2.3M	81.8%	60.5%	5.8 MB	2.1 ms
YOLOv8n	3.2M	82.1%	61.2%	6.5 MB	2.8 ms
YOLOv5n	2.5M	79.2%	58.5%	5.3 MB	2.5 ms
MobileNet + SSD	6.8M	73.8%	49.2%	27 MB	4.5 ms
YOLO11n (COCO pretrained, no VOC fine-tune)	2.6M	35.0%	22.0%	5.5 MB	2.3 ms

Why this model scores higher than full YOLO11n on VOC:

• The backbone was frozen, so it kept strong COCO features.
• Only the neck and head were adapted to the 20 VOC classes.
• This reduces overfitting on the small VOC dataset and keeps the model lightweight.

What It Can Detect

20 classes: aeroplane, bicycle, bird, boat, bottle, bus, car, cat, chair, cow, diningtable, dog, horse, motorbike, person, pottedplant, sheep, sofa, train, tvmonitor.

It supports multi-object detection, up to 300 detections per image.

Usage

from ultralytics import YOLO

model = YOLO("Falcon1-2.6M.pt")

# Image
results = model.predict(source="image.jpg", imgsz=640, conf=0.25)
results[0].show()

# Video
results = model.predict(source="video.mp4", imgsz=640, save=True)

# Validation
metrics = model.val(data="VOC.yaml")
print(metrics.results_dict)

ONNX

For CPU and edge devices without PyTorch, use the ONNX export:

model = YOLO("Falcon1-2.6M.pt")
model.export(format="onnx", imgsz=640, nms=True)

ONNX file is about 10 MB in FP32. It runs with onnxruntime or opencv.dnn.

Training Details

• Base: YOLO11n pretrained on COCO
• Dataset: Pascal VOC 2007 + 2012 (trainval)
• Training: 30 epochs, batch 32, image size 640, freeze=10 (backbone frozen), lr 0.01, amp True
• Hardware: 1x T4
• Total training time: 2.05 hours
• Augmentation: mosaic, mixup off in last 10 epochs, randaugment

Limitations

• Small objects (bottle, pottedplant, chair) have lower accuracy.
• Dark or unusual poses can be confused. For example, a black cat lying on its side in low light can be confused with a dog.
• Trained only on VOC. It does not know COCO classes outside the 20 VOC classes.
• Not trained for video tracking, only per-frame detection.

License

Apache 2.0. Free for personal and commercial use. Includes an explicit patent grant, so contributors cannot sue users over patents covering the model.