adelai-det

# AdelaiDet AdelaiDet is an AIM/Adelaide Detectron2-based research platform for instance-level recognition: object detection, instance segmentation, text spotting, keypoint detection, and related deployment utilities. Use this repo skill when a task names AdelaiDet, `adet`, FCOS, BlendMask, CondInst, BoxInst, SOLOv2, BAText/ABCNet, MEInst, FCPose, DenseCL, or asks how to train, evaluate, demo, prepare data, or export models for this repository. ## Start here - Read `references/repo-provenance.md` before refreshing the skill or checking whether the source snapshot matches a task. - Read `references/compatibility.md` before installing or building AdelaiDet. This repo is legacy Detectron2 code and needs a version-compatible PyTorch/CUDA stack. - Read `references/model-overview.md` to choose a config family and understand which workflow owns it. - Read `references/api-reference.md` for the verified import surface, config keys, registries, custom ops, and public CLIs. - Read `references/troubleshooting.md` when install, import, CUDA extension, CLI, dataset, checkpoint, or export errors appear. ## Install and smoke-check The verified runtime stack is CUDA-capable and legacy-compatible: - Python 3.9 - PyTorch 1.10.x with CUDA 11.3 - TorchVision 0.11.x - Detectron2 0.6 built for the same PyTorch/CUDA pair - AdelaiDet installed editable from a matching source checkout - Pillow `<10`, rapidfuzz `<3`, NumPy `1.23.x`, and OpenCV headless `4.8.x` Do **not** start with a modern PyTorch 2.x stack for unmodified AdelaiDet CUDA extensions: the source includes legacy THC headers in `ml_nms.cu` that are absent from PyTorch 2.x. After installation, run the skill-owned smoke check: ```bash python scripts/check_install.py --cuda-ops ``` Run without `--cuda-ops` only when you intentionally need a CPU/import-only diagnosis. ## Route map ### `setup-build` Use this route for environment creation, Detectron2/PyTorch/CUDA versioning, editable builds, compiled `adet._C` checks, custom op smoke tests, and install failure diagnosis. Read: - `sub-skills/setup-build/SKILL.md` - `sub-skills/setup-build/references/setup-build.md` - `sub-skills/setup-build/references/runtime-checks.md` ### `train-eval` Use this route for Detectron2-style AdelaiDet training, evaluation, config overrides, model-family selection for training, checkpoints, distributed launches, and output directory expectations. Read: - `sub-skills/train-eval/SKILL.md` - `sub-skills/train-eval/references/train-eval-workflows.md` - `sub-skills/train-eval/references/config-selection.md` ### `demo-visualize` Use this route for image/video/webcam demos, `VisualizationDemo`, confidence thresholds, text/non-text visualizations, and dataset visualization. Read: - `sub-skills/demo-visualize/SKILL.md` - `sub-skills/demo-visualize/references/demo-workflows.md` - `sub-skills/demo-visualize/references/visualization.md` ### `text-spotting` Use this route for ABCNet/BAText, BezierAlign, text datasets, custom dictionaries, lexicons, text evaluation, and OCR-specific pitfalls. Read: - `sub-skills/text-spotting/SKILL.md` - `sub-skills/text-spotting/references/text-workflows.md` - `sub-skills/text-spotting/references/text-data-and-eval.md` ### `data-prep` Use this route for COCO/PIC/LVIS/text dataset layouts, semantic mask generation, dataset registration, mapper expectations, MEInst mask encoding, and data validation. Read: - `sub-skills/data-prep/SKILL.md` - `sub-skills/data-prep/references/dataset-preparation.md` - `sub-skills/data-prep/references/data-formats.md` ### `export-convert` Use this route for checkpoint key conversion, optimizer stripping, FCOS/BlendMask weight migration, ONNX export, and optional Caffe/NCNN/TensorRT deployment caveats. Read: - `sub-skills/export-convert/SKILL.md` - `sub-skills/export-convert/references/export-and-checkpoints.md` - `sub-skills/export-convert/references/onnx-export.md` ## Skill-owned scripts - `scripts/check_install.py` — verify import, Detectron2 registries, config keys, and optionally CUDA custom ops. - Sub-skill scripts wrap or adapt the repository workflows with preflight validation. When a script asks for `--repo-root`, pass a source checkout matching the provenance baseline or a refreshed AdelaiDet checkout. ## Operating cautions - Full training, evaluation, demos with real images, and ONNX runtime validation need external datasets, model weights, and sometimes extra runtimes. Use help/dry-run checks first. - ONNX/Caffe/NCNN/TensorRT shell pipelines from the source repository are reference-only here because they assume external workspaces and large artifacts. - Keep installation/build issues routed to `setup-build`; do not debug model configs until `scripts/check_install.py --cuda-ops` passes for CUDA workflows.