How to create Ground Truth and Prediction txt files

[hiteshkhanna31]

trafficstars

I am using Tensor flow 2 for YOLOV3 (https://github.com/pythonlessons/TensorFlow-2.x-YOLOv3) and i want to evaluate it on MS COCO validation dataset. How should the prediction code be modifed so that i have a detection txt generate for every image and how can i make Groundtruth txt from instance_val2017.json files available in the dataset.

I want to use this Object detection metrics to calculate mAP.

Appreciate if someone can help as i am beginner in Computer vision domain.

[takehiro-code]

In my understanding, this repository computes mAP (also called as AP) based on PASCAL VOC2012 challenge. The metric mAP is computed based on the average of all object classes and IOU threshold at 0.5. However, if you want to evaluate based on COCO challenge, mAP will be computed based on the average of all object classes AND average of 10 IOU thresholds of 0.5 to 0.95. I think modification of mAP definition to COCO challenge may be necessary, which I hope as a new feature or enhancement.

[Ash-Chump2Champ]

Hi so I am using that script to convert YOLO to our format: I would like to ask, do I need to use the .cfg file for training or testing?

Testing

batch=1 subdivisions=1

Training

#batch=64 #subdivisions=16