Skip to content

Gradiant/pyodi

Repository files navigation

Pyodi

Python Object Detection Insights

Documentation: https://gradiant.github.io/pyodi

Introduction

A simple tool for explore your object detection dataset. The goal of this library is to provide simple and intuitive visualizations from your dataset and automatically find the best parameters for generating a specific grid of anchors that can fit you data characteristics

Component Description
paint annotations paints COCO format annotations and predictions
ground-truth explore your dataset ground truth characteristics
evaluation evaluates AP and AR for between predictions and ground truth
train-config generation automatically generate anchors for your data
train-config evaluation evaluate the fitness between you data and your anchors
coco merge automatically merge COCO annotation files
coco split automatically split COCO annotation files in train and val subsets
crops split creates a new dataset by splitting images into crops and adapting the annotations file
crops merge translate COCO ground truth or COCO predictions crops split into original image coordinates

Installation

pip install pyodi

Usage

Pyodi includes different applications that can help you to extract the most from your dataset. You can download our TINY_COCO_ANIMAL dataset here in order to test the example commands. A classic flow could follow the following steps:

1. Annotation visualization

With pyodi paint_annotations you can easily visualize in a beautiful format your object detection dataset.

pyodi paint-annotations \
  $TINY_COCO_ANIMAL/annotations/train.json \
  $TINY_COCO_ANIMAL/sample_images \
  $TINY_COCO_ANIMAL/painted_images

COCO image with painted annotations

2. Ground truth exploration

It is very recommended to intensively explore your dataset before starting training. The analysis of your images and annotations will allow you to optimize aspects as the optimum image input size for your network or the shape distribution of the bounding boxes. You can use the ground_truth app for this task:

pyodi ground-truth $TINY_COCO_ANIMAL/annotations/train.json

Image shape distribution

Bbox distribution

Bbox center distribution

3. Train config generation

The design of anchors is critical for the performance of one-stage detectors. Pyodi can help you to automatically design a set of anchors that fit your data distribution.

pyodi train-config generation \
  $TINY_COCO_ANIMAL/annotations/train.json \
  --input-size [1280,720] \
  --n-ratios 3 --n-scales 3

Anchor clustering plot

4. Train config evaluation

Pyodi evaluation app has been designed with the aim of providing a simple tool to understand how well are your anchors matching your dataset. It automatically runs by default after executing train-config generation but it can also be run independently with:

pyodi train-config evaluation \
  $TINY_COCO_ANIMAL/annotations/train.json \
  $TINY_COCO_ANIMAL/resources/anchor_config.py \
  --input-size [1280,720]

Anchor overlap plot

Contributing

We appreciate all contributions to improve Pyodi. Please refer to Contributing guide for more info.