Task examples¶

The repository ships two runnable scripts per task family under examples/tasks/:

<task>.py: inference only — load weights → predict → annotate, no training;
<task>_train.py: training example, which generates a small synthetic dataset by default (detection uses coco8).

The examples live outside src/autovision/tasks/: they are not shipped in the wheel, and training examples import the CLI orchestration layer, which would break the repository's one-way layering contract.

Inference examples:

default image: examples/tasks/assets/airport_aerial.jpg (CC0, San Diego airport aerial view; source and sha256 in assets/README.md);
support --weights (an exported .pt), --image and --threshold;
write prediction.jpg + detections.json.

Training examples:

run 1–3 steps with a tiny model configuration; CPU is enough for most tasks
default to the GTR family: run 1–3 steps with a tiny random-init GTR configuration; this needs .[gtr] + CUDA (only panoptic/classification have no GTR route and stay CPU-friendly);
print the checkpoint, exported model, validation metrics and class map;
support --predict-image, --export and --dry-run.

Install and quick start¶

pip install -e ".[gtr,download,export]"   # GTR examples (CUDA + Triton)

# Inference with official weights, no training
python examples/tasks/detection.py
python examples/tasks/oriented_detection.py
python examples/tasks/instance_segmentation.py
python examples/tasks/semantic_segmentation.py
python examples/tasks/pose_estimation.py
python examples/tasks/depth_estimation.py

# Panoptic/classification have no official task weights; pass an exported .pt
python examples/tasks/panoptic_segmentation.py \
    --weights runs/examples/panoptic_segmentation_train/exported_models/exported_last.pt

# Training examples
python examples/tasks/detection_train.py --steps 2
python examples/tasks/classification_train.py --steps 2
python examples/tasks/instance_segmentation_train.py --dry-run

All inference examples load the official GTR MIT weights by default: detection detr/gtrvit-s, and the other tasks gtrobb_s_dota / gtrseg_s_coco / gtrsemseg_s_cityscapes / gtrpose_s_coco / gtrdepth_s (detection accepts --model rfdetr/rf-detr-nano for a CPU-friendly released checkpoint). All inference scripts default to the CC0 assets/airport_aerial.jpg (its overhead domain matches the DOTA-trained OBB checkpoint); pass --image for street-level or task-specific imagery. GTR requires:

pip install -e ".[gtr]"   # CUDA + Triton

detection.py --effect demonstrates the enhanced annotators migrated from supervision: corner / round replace plain boxes with corner brackets or rounded boxes, while blur / pixelate apply privacy filtering before drawing boxes and labels. The default none keeps the original rendering:

python examples/tasks/detection.py --effect corner
python examples/tasks/detection.py --effect blur --threshold 0.5

YOLO baseline examples (optional, AGPL)¶

examples/tasks/yolo/ adds five Ultralytics YOLO inference examples: detection, segmentation, pose, OBB and classification, reusing the same images. ultralytics is AGPL-3.0 and is not part of the Apache-2.0 core:

pip install -e ".[ultralytics]"
python examples/tasks/yolo/detection.py
python examples/tasks/yolo/segmentation.py
python examples/tasks/yolo/pose.py
python examples/tasks/yolo/obb.py            # airport aerial image
python examples/tasks/yolo/classification.py

Weights are cached under ~/.cache/autovision/yolo, never in the repository. Detection / segmentation / pose weights are downloaded from the ModelScope Ultralytics/YOLO11 mirror; OBB / classification are not in that repository, so they fall back to the Ultralytics default source or a local --model /path/to/*.pt. See examples/tasks/yolo/README.md for details.

Tasks and default weights¶

Task	Inference	Training	Default inference weights
Detection	`detection.py`	`detection_train.py`	`detr/gtrvit-s` (GTR, COCO)
Oriented detection	`oriented_detection.py`	`oriented_detection_train.py`	`gtrobb_s_dota` (GTR, DOTA)
Instance segmentation	`instance_segmentation.py`	`instance_segmentation_train.py`	`gtrseg_s_coco` (GTR, COCO)
Panoptic segmentation	`panoptic_segmentation.py`	`panoptic_segmentation_train.py`	requires `--weights`
Semantic segmentation	`semantic_segmentation.py`	`semantic_segmentation_train.py`	`gtrsemseg_s_cityscapes`
Pose	`pose_estimation.py`	`pose_estimation_train.py`	`gtrpose_s_coco` (GTR, COCO)
Depth	`depth_estimation.py`	`depth_estimation_train.py`	`gtrdepth_s` (GTR)
Classification	`classification.py`	`classification_train.py`	requires `--weights`

The training data contracts are unchanged: detection COCO/YOLO, oriented detection YOLO-OBB, instance COCO instances, panoptic masks + segments, semantic paired images/masks, pose COCO person_keypoints, depth images + depths, and classification ImageFolder. The six GTR routes (detection / OBB / instance / semantic / pose / depth) train tiny random-init configurations by default; official weights finetune via model_args={"checkpoint": "<asset name>"}.

Real-data configuration examples:

# Instance segmentation
data = {"train": {"annotations": "instances_train.json", "images": "train2017"},
        "val":   {"annotations": "instances_val.json",   "images": "val2017"}}

# Semantic segmentation
data = {"train": {"images": "images/train", "masks": "masks/train"},
        "val":   {"images": "images/val", "masks": "masks/val"}}

# Pose
data = {"train": {"annotations": "person_keypoints_train2017.json",
                  "images": "train2017"},
        "val":   {"annotations": "person_keypoints_val2017.json",
                  "images": "val2017"}}

# Depth
data = {"train": {"images": "images/train", "depths": "depths/train"},
        "val":   {"images": "images/val", "depths": "depths/val"}}

See examples/tasks/README.md for the full data schemas, common flags and export notes.

Notes¶

Training examples run only 1–3 steps by default; their metrics are meaningless and only prove the pipeline runs. Use the recipes in benchmarks for real numbers.
--export only copies .pt; use autovision-cli export for ONNX / OpenVINO / TensorRT (numerical parity gates are enabled by default).
Outputs default to runs/examples/<task>/ (inference) or runs/examples/<task>_train/ (training); demo data goes to runs/examples/_demo_data/<task>/ and is not removed by --overwrite.