Ctrl+K

WorkShop - Object Detection with Yolov8

WorkShop - Object Detection with Yolov8#

env#

  • venv

micromamba create -n mambaYoloWrkShp -f EnvObjectDetection.yml
micromamba activate mambaYoloWrkShp

dataset#

labels#

  • labels standards:

    COCO Style: [x, y, width, height].
    PASCAL VOC Style: [x_min, y_min, x_max, y_max].
    YOLO Style: [x_center, y_center, width, height].

model#

Guidelines:

  • The 2 Stages models and ViT based models are the most accurate, yet usually slowest.

  • The YOLO based models are the fastest. Commonly used in edge devices or where real time performance are crucial.

  • The SSD based models are usually a balance and ranked in the middle with respect to speed and accuracy.

  • yolo main concept is based on a Grid + Anchors.

  • There are variants of YOLO which are anchor free for example: https://scribe.rip/fc0e25622e1c

alt text

resources#

arch V8#

Train#

  • Choose the Model

  • Choose the Head - Modern YOLO models can do object detection, segmentation and key points detection.

  • Select the Hyper Parameters

  • train

score#

alt text

alt text

mAP alternative measures:

  • Performance on specific type of objects (small objects)

  • Analyze classification errors (Confusion Matrix).

  • Analyze regression errors: Size, Locations (Biases).

  • Analysis per class / weighted by classes

lab#

  • test Model: Ultralytics YOLOv8n- Set for Object Detection.

    • (n) is due to the simplicity of the task and the constrained time for training.

    • Image Input Size: 640x640.

    • Number of Parameters: 3.2e6.

  • input image is relatively large (1920x1080)

    • The model was trained on 640x640 images with [80x80, 40x40, 20x20] grids.

    • Naive training means the data will be resized with factor 3.

    • The ball size (Bounding Box) is about ~12x12.

    • After factorization it will be ~4x4 pixels which is too low for valid performance.

  • Solution: Tiled Detection (obss/sahi)

0001BoundingBoxFormat.ipynb#

  • convert between types

0002Dataset.py#

  • dw data set

0003PreProcessTiles.py#

  • prepare tiles

Train/Tiles/Video001Frame0015991Tile005.png
Video001Frame0015991Tile003.png  Video001Frame0015991Tile004.png  Video001Frame0015991Tile005.png  Video001Frame0015991Tile008.png  Video001Frame0015991Tile009.png  Video001Frame0015991Tile010.png

alt text

alt text

0005TrainYolo.py#

from ultralytics import YOLO


modelCfgFile        = 'yolov8n.yaml' #<! The name sets the scale of the model
modelWeightsFile    = 'yolov8n.pt' #<! Download from GitHub


modelYolo = YOLO(modelCfgFile, task = 'detect')
modelYolo = modelYolo.load(modelWeightsFile)


modelResults = modelYolo.train(data = dataFile, epochs = numEpoch, batch = batchSize, imgsz = imgSize, workers = numWorkers, name = 'DetectionRun', amp = ampMode)
Contents