No accounts, no telemetry, no cloud
Storage is a SQLite DB plus plain folders under data/. Delete the folder and the data is gone — nothing is phoned home.
Binds to 127.0.0.1
The web UI is not reachable from the network by default. Put a reverse proxy in front if you deliberately want LAN access.
Air-gapped capable
The only outbound traffic is downloading model weights the first time. Supply .pt files yourself and it never calls out.
Run YAAP in two minutes
Pull the pre-built image from Docker Hub, or clone the repo and build it yourself. Both give you the identical app at http://127.0.0.1:8811, with all state persisted on your own disk.
From GitHub
Clone the repo and build locally with Docker Compose — GPU by default, CPU fallback profile included.
# clone git clone https://github.com/Computer-Vision-AI-Hub/YAAP cd YAAP # GPU (needs NVIDIA driver + nvidia-container-toolkit) docker compose up --build # CPU fallback docker compose --profile cpu up --build yaap-cpu
http://127.0.0.1:8811 once it's up. Full setup notes in the README.From Docker Hub
Skip the build — pull the published image straight from Docker Hub and run it.
# GPU image
docker pull lukasiktar/yaap:gpu
docker run -d --name yaap --gpus all \
-p 127.0.0.1:8811:8811 \
-v $(pwd)/data:/app/data \
-v $(pwd)/seg_models:/app/seg_models \
--shm-size=8gb \
lukasiktar/yaap:gpu
# CPU image (no GPU on this machine)
docker pull lukasiktar/yaap:cpu
docker run -d --name yaap-cpu \
-p 127.0.0.1:8811:8811 \
-v $(pwd)/data:/app/data \
-v $(pwd)/seg_models:/app/seg_models \
lukasiktar/yaap:cpu
Everything from first click to trained model
One interaction model, three task types — object detection, instance segmentation, and classification — each exported in the format its architecture actually trains on.
Tasks
Object detection (boxes), instance segmentation (polygons), classification.
Editor
Pan/zoom canvas, SAM assist, box + polygon tools, vertex editing, class hotkeys 1–9, autosave.
SAM assist
SAM 3 point/box prompts, mask→polygon simplification, task-aware conversion, per-mask class picker.
Auto-label
Batch inference with YOLOv8/11/26, RT-DETR, or your own trained weights. Predictions land flagged for review.
Augmentation
albumentations pipeline — flips, rotate, blur, noise, CLAHE, cutout and more — with live preview.
Versions
Frozen dataset snapshots: split ratios, preprocessing, augmentation multiplier, zip download.
Export
YOLO (detect / seg / cls, data.yaml) or COCO JSON, Roboflow-style layout.
Training
Launch ultralytics training jobs from the UI with a live log. Best weights feed back into auto-label.
GPU
PyTorch + CUDA, auto-detected. SAM, auto-label and training all use it when available.
SAM-assisted annotation is the heart of YAAP
Click or drag on the object — SAM 3 predicts a mask instantly. Refine it with more clicks, then pick a class from the chip row or hit Accept. YAAP converts the mask to whatever format your project trains on.
Detection project → enclosing box. Segmentation project → editable polygon. One interaction model, correctly-formatted labels either way.
| Prompt | How |
|---|---|
| Point | Click on the object (shift-click adds a negative point to carve away background). Every extra click refines the mask. |
| Box | Drag a rough rectangle around the object. |
| Reject | Esc discards the predicted mask with no trace left behind. |
SAM 3 weights are gated by Meta — they don't auto-download like the
YOLO/RT-DETR weights do. Request access on the
SAM 3 Hugging Face page,
wait for approval, then authenticate with hf auth login before first use
(or download the checkpoint yourself once approved and drop it into seg_models/).
From upload to a trained model
A human-in-the-loop flywheel: label a little, auto-label the rest, review, retrain, repeat.
New project
Pick the task type — it decides the tools you see and the export format.
Classes
Define them before labeling; their order becomes the YOLO class index.
Upload
Drag & drop — originals stored untouched, thumbnails generated automatically.
Annotate
S SAM assist · B box · P polygon · 1-9 class · saves automatically.
Auto-label
Run a pretrained / uploaded / YAAP-trained model over unannotated images, then review.
Generate
Choose splits, preprocessing and augmentations, preview them, generate a version.
Versions
Download the zip, or point yolo train straight at the version folder.
Train
Pick version + architecture + epochs, watch the live log. best.pt feeds the next auto-label round.
See it in action
Gallery is on its way — drop images into docs/img/ and swap these placeholders for real captures.
More to come as the platform gets used in real projects.
Built by ComputerVisionAIHub
YAAP is part of ComputerVisionAIHub's open-source computer-vision tooling. Check out the hub for more projects, or come back here once images and case studies are ready to add.
Visit ComputerVisionAIHub →