Object detection

[4F2E4A2E]

Hi there and thank you for piwise!

My question is: how could piwise or any segnet implementation be used for object detection?

[bodokaiser]

Depends on the object(s)?

You could train on your own dataset with custom classes wherein every class corresponds to the object you want to detect.

This however require you to have the labeled data.

Am 15.10.2017 um 21:51 schrieb 4F2E4A2E notifications@github.com:

Hi there and thank you for piwise!

My question is: how could piwise or any segnet implementation be used for object detection?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/bodokaiser/piwise/issues/12, or mute the thread https://github.com/notifications/unsubscribe-auth/ABsq8kwO186XsgdK8TDJGxuNBJ9zJrc-ks5ssmJMgaJpZM4P53ni.

[4F2E4A2E]

Can you please explain on how it would depend on the object?

[bodokaiser]

From what should the network know if you want to detect an apple or a car?

Am 15.10.2017 um 21:58 schrieb 4F2E4A2E notifications@github.com:

Can you please explain on how it would depend on the object?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/bodokaiser/piwise/issues/12#issuecomment-336737173, or mute the thread https://github.com/notifications/unsubscribe-auth/ABsq8hMzKgEhCptMJ5rECdux3Ctj8bmcks5ssmPkgaJpZM4P53ni.

[4F2E4A2E]

Please elaborate, could any network know any difference besides the information given by the semantic segmentation labeling?

[bodokaiser]

You can’t say for sure what happens in the hidden layers but generally the network will only output what it was trained with.

Am 15.10.2017 um 22:03 schrieb 4F2E4A2E notifications@github.com:

Please elaborate, could any network know any difference besides the information given by the semantic segmentation labeling?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/bodokaiser/piwise/issues/12#issuecomment-336737505, or mute the thread https://github.com/notifications/unsubscribe-auth/ABsq8rpwevk2Yam2gLXE40pf3w3lZleUks5ssmUtgaJpZM4P53ni.

[4F2E4A2E]

Exactly, that's my honest opinion too, but my question is, how to train with the segmented ground truth data?

[bodokaiser]

You should checkout „fine-tuning“.

Am 15.10.2017 um 23:05 schrieb 4F2E4A2E notifications@github.com:

Exactly, that's my honest opinion too, but my question is, how to train with the segmented ground truth data?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/bodokaiser/piwise/issues/12#issuecomment-336741599, or mute the thread https://github.com/notifications/unsubscribe-auth/ABsq8l1Dv6x0zwVSzd_6gO_Pf3y1zmlIks5ssnOogaJpZM4P53ni.

[4F2E4A2E]

This one? https://arxiv.org/pdf/1601.05150.pdf

[bodokaiser]

Yes that is one application. Actually „fine-tuning“ is quite a common term in deep learning:

http://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html

It is also required by some of the semantic segmentation networks which use a vgg trained network and fine tune it to the PASCAL VOC class labels.

Am 16.10.2017 um 00:31 schrieb 4F2E4A2E notifications@github.com:

This one? https://arxiv.org/pdf/1601.05150.pdf https://arxiv.org/pdf/1601.05150.pdf — You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/bodokaiser/piwise/issues/12#issuecomment-336746591, or mute the thread https://github.com/notifications/unsubscribe-auth/ABsq8vzznW8z7AKai1dM6-F_L2lITs0rks5ssofFgaJpZM4P53ni.