Nobunaga is an Object Detection Analysis tool in Computer Vision. Based on TIDE, which is one of the excellent error analysis algorithms, Nobunaga exports the overall distribution of detection error in the given dataset as well as per-category error. Towards analyzing miss classification, Nobunaga plots their error in a confusion matrix. Moreover, it plots detection errors in images one by one with annotation.
In addition, Nobunaga can calculate Optimal Correction Cost (OCC) per image. The numeral score proposed by Otani et al. indicates how much it costs to bring a prediction closer to the corresponding ground truth. Nobunaga allows the users to see how wrong a detection was, image by image and label by label. We hope it could be beneficial for you to figure out what kind of detection error happens in your models specifically.
pip install nobunaga
or
git clone https://github.com/FastAccounting/nobunaga
cd nobunaga
pip install -e .
cd nobunaga
nobunaga
Before analyzing detection error, you need to prepare ground truth formated in COCO object detection and its corresponding prediction results in JSON and image dataset.
Command line is as follows:
nobunaga --pred coco_instances_results.json \
--gt instances_val.json \
--image_dir path/to/image_dir \
--iou_threshold 0.5 \
--confidence_threshold 0.7 \
--output_image
Then, you can get the below files:
- Error summary similar to TIDE
- Per-label detection error
- Class error matrix
- Error images with ground truth label
- Error scores of each image based on OCC (Details below.)
The definition of errors follows TIDE. However, unlike TIDE, we remark that Nobunaga counts the number of labels for each error. Therefore, Nobunaga displays the total number of corresponding errors in the data set. The numbers are different from those in TIDE.
Outputs related to OCC are generated under the _occ folder when Nobunaga is run. The contents are the following three items:
- A summary of each image about OCC in CSV.
- Per-label OCC as summarized in a CSV file.
- Directories of images containing the six TIDE error types were sorted into folders based on OCC score.
The summary of errors provides an overview of errors in the dataset, similar to TIDE.
This section delves into the Error Summary and shows which errors occur for each label in a figure. The further to the right, the higher the number of errors. For example, we confirm that the "Person" label has many Dupe errors, followed by Loc, Bkg, and Miss.
The confusion matrix visualizes what label tends to be misclassified as other labels. This confusion matrix visualizes only false examples, so correct ones are omitted.
Images containing detection errors are written to the {model_name}_error
directory for each error.
The pictures show one with the error on the left and another with the ground truth on the right.
Two examples are shown here.
The first image is an example of a Bkg error, which incorrectly detects the background area. The correct image has no annotations.
The second image is an example of a Miss error, indicating a detection error has been occured.
localized correctly but classified incorrectly. max(IoU) ≥ tf for GT of the incorrect class.
Detected by the small bounding box below IoU threshold. Classified correctly but localized incorrectly. tb ≤ max(IoU) ≤ tf for GT of the correct class
Classified incorrectly and localized incorrectly. tb ≤ max(IoU) ≤ tf for GT of the incorrect class
It would be correct if not for a higher scoring detection. max(IoU) ≥ tf for GT of the correct class but another higher-scoring detection already matched that GT
Detected background as foreground. max(IoU) ≤ tb for all GT.
All undetected ground truths (false negatives) are not already covered by classification or localization error. In Nobunaga, "not already covered by all other errors" define as Miss Error, and more labels tend to be explained as Miss Error more than other errors.
We use GenEi Gothic-Font to visualize above errors.
Distribution site