Train a custom object detection model for LaiCai Flow

• Include different positions, sizes, lighting, backgrounds, animation frames, and partial occlusion.
• Keep difficult examples: tiny objects, edge objects, blur, popups, and visually similar non-targets.
• Add some background images where none of your target classes appear. Their label files are empty.
• There is no universal minimum. For a first test, aim for at least 50 varied boxes per class; difficult or similar classes usually need 100–200 or more.

• Every visible instance of every target class must be labeled. An unlabeled target is learned as background.
• Use the same box policy everywhere: tight around the visible object, without large margins.
• Prefer names such as CopperDepleted or copper_depleted when a labeling tool changes spaces.

1. 1

   Open MakeSense.ai and upload the source images. Keep your original image folder; the exported label ZIP does not replace it.

2. 2

   Choose Object Detection, not Image Recognition.

3. 3

   Create the final class list in a fixed order, then draw a rectangle around every target instance.

4. 4

   Review empty images, missed objects, wrong classes, and loose boxes before export.

5. 5

   Export Rect annotations as a ZIP package in YOLO format.

dataset/
├── images/
│   ├── train/
│   └── val/
├── labels/
│   ├── train/
│   └── val/
└── data.yaml

path: /absolute/path/to/dataset
train: images/train
val: images/val

names:
  0: Copper
  1: CopperDepleted
  2: Wood
  3: Monster

A practical starting split is 80% train and 20% validation. The validation set must contain genuinely unseen scenes.

• For a new annotation batch, you do not need to upload the old images to MakeSense.ai again.
• Merge the new images and labels offline, then remap numeric class IDs by class name so the original order stays unchanged.
• Retrain with the complete old + new dataset. Training only on the new images can make old classes behave like background.

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install ultralytics==8.4.138 onnx==1.22.0 \
  onnxruntime==1.29.0 onnxslim==0.1.96

Recommended starting parameters

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.train(
    data="dataset/data.yaml",
    imgsz=960,
    epochs=200,
    patience=40,
    batch=8,
    device="mps",  # NVIDIA: 0 · CPU: "cpu"
    workers=4,
    seed=42,
    close_mosaic=15,
)

• Precision: how many reported detections are correct.
• Recall: how many real objects were found.
• mAP50 and mAP50-95: overall localization and classification quality across thresholds.
• Confusion matrix: which classes are mistaken for each other or for background.

from ultralytics import YOLO

model = YOLO("runs/detect/train/weights/best.pt")
model.export(
    format="onnx",
    imgsz=960,
    batch=1,
    dynamic=False,
    simplify=True,
    nms=None,
)

Package the model directory

Select the directory itself in LaiCai's Model library. The directory must contain config.json and the ONNX file named by modelFile. The class IDs must start at 0 and remain contiguous.

game-objects/
├── config.json
└── model.onnx

{
  "id": "game-objects-v1",
  "name": "Game Objects v1",
  "type": "yolo",
  "modelFile": "model.onnx",
  "outputFormat": "raw_yolo_no_objectness",
  "id2label": {
    "0": "Copper",
    "1": "CopperDepleted",
    "2": "Wood",
    "3": "Monster"
  },
  "label2id": {
    "Copper": 0,
    "CopperDepleted": 1,
    "Wood": 2,
    "Monster": 3
  }
}

• Select the model in Model library and confirm its input, output, output format, and class list.
• Use Test Current Screen with a class and a reasonable starting threshold such as 0.5.
• Test both positive frames and negative or confusing frames. A successful inference call is not the same as a detected object.
• After desktop validation, test the same model separately on Android Agent if the Flow will run on the phone.