Object Detection Metrics and Analysis in Python

Creating evaluation metrics for projects involving object detection takes a surprising amount of time. This repo contains code we've found useful to speed up the results analysis for object detection projects. It provides:

1. Easy creation of a pandas inference dataframe enabling detailed analysis.
2. Summary statistics for easy plotting.
3. Calculation of coco metrics using the same pandas dataframe. (uses pycocotools).

To see a quick example of the functionality have a look at the starter notebook.

Installation

pip install git+https://github.com/alexhock/object-detection-metrics

Imports:

from objdetecteval.metrics import (
    image_metrics as im,
    coco_metrics as cm
)

Example usage

Take predictions in a pandas dataframe and similar labels dataframe (same columns except for score) and calculate an 'inference' dataframe:

infer_df = im.get_inference_metrics_from_df(preds_df, labels_df)
infer_df.head()

The inference dataframe enables easy analysis of the results for example:

1. IoU stats by class and failure category
2. Highest scoring false positive predictions
3. Comparison of bounding box distributions for FP and TP
4. ... etc. ..

class_summary_df = im.summarise_inference_metrics(infer_df)
class_summary_df

This makes it easy to plot:

figsize = (5, 5)
fontsize = 16

fig_confusion = (
    class_summary_df[["TP", "FP", "FN"]]
    .plot(kind="bar", figsize=figsize, width=1, align="center", fontsize=fontsize)
    .get_figure()
)

fig_pr = (
    class_summary_df[["Precision", "Recall"]]
    .plot(kind="bar", figsize=figsize, width=1, align="center", fontsize=fontsize)
    .get_figure()
)

Coco metrics are just as simple

Use the dataframes to calculate full coco metrics

res = cm.get_coco_from_dfs(preds_df, labels_df, False)
res