hustqbw / single-object-localization

Object Localization via CNN

with anchor method

The .py file with name "with_anchor" is the algorithm which uses anchor to let the model learn the offset of bounding boxes.

You can compare them with the original algorithm and codes without anchors.

Task and Dataset Info

• Object localization using CNN
• Data: In each image, there is just one object. We just need to localize the only object in the image. If you are interested in multi-object detection(object detection field),you can modify my code to reach your goal.
• Evaluation metric: A correct localization means the predict class label is correct and the predict box has an loU>0.5 with the ground- truth.
• implementing the localization network using direct bbox regression

Because this is an experiment of 《Foundation of Computer Vision》class which is hosted by Prof.Xinggang Wang and Prof.Zhenyu Liao in Huazhong University of Science and Technology. So I am not sure whether I have the privilege to make the dataset open source. Whatever, if there is any question, please contact on me! I will pull it at soon.

dataset is shown in the .zip file

some dataset info

data: image with only one object inside

label: the class info

bbox: the (xmin,ymin,xmax,ymax) bounding box coordinate

Result