aerogaze-yolo26l

AeroGaze YOLO26-L L0 E50

This repository contains the final detector checkpoint used by the AeroGaze active-camera project. The detector is a YOLO26-L model fine-tuned for 16 aerial-object classes. Camera control, Kalman memory, tests, and deployment code are available in the AeroGaze GitHub repository.

Model details

• Architecture: YOLO26-L object detector
• Framework: Ultralytics 8.4.155 and PyTorch 2.9.1
• Input used by the deployment pipeline: 1280 pixels with aspect-ratio-preserving rectangular preprocessing
• Runtime precision used during deployment: FP16 on CUDA
• Output: source-frame normalized xyxy boxes, class labels, and confidence scores
• Training stage represented by this checkpoint: 50 cumulative epochs
• Final stage: 10 low-learning-rate epochs from the E40 checkpoint
• Final-stage optimizer: AdamW with learning rate 1e-5
• Final-stage images: 47 L0 views
• Final-stage optimizer updates: 120

The checkpoint contains no optimizer state. Local filesystem paths embedded by the training framework were removed before publication without changing any model tensor.

Published checkpoint SHA256: 21d84ae3f67967ac08190a1742e6b4e2c95160642db4039ce9f3ad68212b29d5.

Classes

The class order is fixed:

hangar
helicopter
jet_plane
large_launcher
large_tower
medium_launcher
medium_plane
mine_roller
small_launcher
small_plane
small_tower
ta-ta
tank
condor
jammer
spacecraft

Usage

Install a CUDA-compatible PyTorch build and Ultralytics 8.4.155, then load the checkpoint:

from ultralytics import YOLO

model = YOLO("model.pt")
results = model.predict("image.png", imgsz=1280, rect=True, conf=0.001)

The AeroGaze deployment pipeline applies original, horizontal-flip, vertical-flip, and combined-flip inference, preserves the top two class alternatives, and fuses boxes. Loading this checkpoint alone does not reproduce that complete pipeline.

Evaluation status

This exact E50 detector with the final active-camera policy does not have a confirmed complete official validation score. A different historical E40 fixed-L0 baseline scored 0.4950687646541436 on the competition validation service and must not be attributed to this checkpoint.

The final camera policy passed a causal offline replay over 248 saved frames, including 50 synthetic L1 observations. Those L1 images were interpolated crops from saved 960x540 frames, not true high-resolution observations, so the replay is a behavior test rather than an accuracy benchmark.

Limitations

• The 47 final-stage images include earlier training and development images. They do not provide an independent held-out estimate.
• The model is specialized for the competition camera geometry and class set.
• Small objects, partial objects, domain shift, compression, and unseen backgrounds may reduce accuracy.
• Confidence values are not calibrated probabilities.
• The checkpoint alone does not include tracking, memory, or camera control.
• Training data and evaluation images are not distributed with this model.

License

The checkpoint is distributed under AGPL-3.0 to match the Ultralytics training stack and pretrained-model license. The separate AeroGaze source repository is MIT licensed; dependencies and model weights retain their own licenses.