search

satojkovic / DeepLogo

A brand logo detection system using tensorflow object detection API.
MIT License
512 stars 157 forks source link

Precision/recall #36

Open jbogp opened 4 years ago

jbogp commented 4 years ago

Hello, I'm currently iterating over DeepLogo with some more recent object detection models, and was wondering if you could share the global average precision/recall you obtained on the flickr logos 27 dataset after training DeepLogo?

satojkovic commented 4 years ago

Hi @jbogp

I've evaluated DeepLogo using mean average precision because Tensorflow Object Detection API provides an evaluation script for this metric. Do you know any publicly available script for global average preison/recall metric?

jbogp commented 4 years ago

Hi @satojkovic I'm sorry, I wrote too fast I meant mean average precision.

satojkovic commented 4 years ago

@jbogp

Never mind. Here is my evaluation result. PascalBoxes_Precision/mAP@0.5IOU: 0.858025

If it's possible, could you share your evaluation results when you finished training?

Thanks!

jbogp commented 4 years ago

Hi @satojkovic Sorry for the delayed reply. Cool, so I got DetectionBoxes_Precision/mAP@.50IOU: 0.8595 and DetectionBoxes_Precision/mAP@.75IOU:0.725 using faster_rcnn_inception_resnet

On the 0.5IOU it's pretty similar so I guess SSD which I think is simpler would be the better choice.

satojkovic commented 4 years ago

Hi @jbogp

Very interesting results! It might be a good idea to train with larger dataset than flickr_logos_27 when you use bigger models.