ODAch, An Object Detection TTA tool for Pytorch

ODA is a test-time-augmentation (TTA) tool for 2d object detectors.

For use in Kaggle object detection competitions.

⭐ if it helps you! ;)

YOLO Integration

ODAch supports YOLOv5, YOLOv8, YOLOv9, YOLOv10, and YOLOv11 models from Ultralytics!

Quick Start with YOLO

import odach as oda
from ultralytics import YOLO

# Load your YOLO model
model = YOLO('yolov8n.pt')  # or yolov5, yolov9, yolov10, yolov11

# Wrap the YOLO model for ODAch
yolo_wrapper = oda.wrap_yolo(model, imsize=640, score_threshold=0.25)

# Define TTA transformations
tta = [oda.HorizontalFlip(), oda.VerticalFlip(), oda.Rotate90Left(), oda.Rotate90Right()]

# Create TTA wrapper
tta_model = oda.TTAWrapper(yolo_wrapper, tta)

# Run inference with TTA
results = tta_model(images)

YOLO Features

• Multi-version support: YOLOv5, YOLOv8, YOLOv9, YOLOv10, YOLOv11
• Automatic format conversion: Handles YOLO output format automatically
• Batch processing: Process multiple images efficiently
• Configurable thresholds: Adjust confidence and IoU thresholds
• Seamless integration: Works with existing ODAch TTA pipeline

YOLO TTA Example

# Advanced YOLO TTA with multiple scales
tta = [
    oda.HorizontalFlip(),
    oda.VerticalFlip(),
    oda.Rotate90Left(),
    oda.Rotate90Right(),
    oda.Multiply(0.9),
    oda.Multiply(1.1)
]

# Multi-scale TTA
scale = [0.8, 0.9, 1.0, 1.1, 1.2]

# Create TTA wrapper with scales
tta_model = oda.TTAWrapper(yolo_wrapper, tta, scale)

# Run inference
results = tta_model(images)

See example_yolo_usage.py and YOLO_INTEGRATION_README.md for detailed examples.

Box Format Support

ODAch supports both VOC and COCO box formats:

Using Different Box Formats

import odach as oda

# VOC format (default): [x1, y1, x2, y2]
tta_model = oda.TTAWrapper(model, tta, box_format="voc")

# COCO format: [x, y, width, height]
tta_model = oda.TTAWrapper(model, tta, box_format="coco")

# Also available for wrap_yolo
yolo_wrapper = oda.wrap_yolo(model, box_format="coco")

Format Conversion Utilities

import odach as oda

# Convert VOC to COCO
coco_boxes = oda.voc_to_coco(voc_boxes)

# Convert COCO to VOC
voc_boxes = oda.coco_to_voc(coco_boxes)

Install

pip install odach

Development Installation

# Using uv (recommended)
uv sync --extra dev

# Run tests
uv run pytest tests/ -v

Usage

See Example.ipynb.

The setup is very simple, similar to ttach.

Singlescale TTA

import odach as oda
# Declare TTA variations
tta = [oda.HorizontalFlip(), oda.VerticalFlip(), oda.Rotate90Left(), oda.Multiply(0.9), oda.Multiply(1.1)]

# load image
img = loadimg(impath)
# wrap model and tta
tta_model = oda.TTAWrapper(model, tta)
# Execute TTA!
boxes, scores, labels = tta_model(img)

Multiscale TTA

import odach as oda
# Declare TTA variations
tta = [oda.HorizontalFlip(), oda.VerticalFlip(), oda.Rotate90Left(), oda.Multiply(0.9), oda.Multiply(1.1)]
# Declare scales to tta
scale = [0.8, 0.9, 1, 1.1, 1.2]

# load image
img = loadimg(impath)
# wrap model and tta
tta_model = oda.TTAWrapper(model, tta, scale)
# Execute TTA!
boxes, scores, labels = tta_model(img)

• The boxes are also filtered by nms(wbf default).

• The image size should be square.

Model Output Wrapping

• Wrap your detection model so that the output is similar to torchvision frcnn format: [{"boxes": [[x,y,x2,y2], ...], "labels": [0,1,..], "scores": [1.0, 0.8, ..]}]

• Example for EfficientDets https://www.kaggle.com/kyoshioka47/example-of-2d-single-scale-tta-with-odach/

# wrap effdet
oda_effdet = oda.wrap_effdet(effdet)
# Declare TTA variations
tta = [oda.HorizontalFlip(), oda.VerticalFlip(), oda.Rotate90Left()]
# Declare scales to tta
scale = [1]
# wrap model and tta
tta_model = oda.TTAWrapper(oda_effdet, tta, scale)

Examples

YOLO TTA Examples

• example_yolo_usage.py - Basic YOLO integration
• YOLO_INTEGRATION_README.md - Detailed YOLO usage guide

Global Wheat Detection

Example notebook

Thanks

nms, wbf are from https://kaggle.com/zfturbo

tta is based on https://github.com/qubvel/ttach, https://github.com/andrewekhalel/edafa/tree/master/edafa and https://www.kaggle.com/shonenkov/wbf-over-tta-single-model-efficientdet