Training code

[wei-m-teh]

Hi @alessiocanepa , I came across your project on github after reading the paper: https://www.mdpi.com/1424-8220/21/4/1252/pdf and was very interested in diving deeper into this algorithm. I have a similar use case that involves tracking a relatively small object (about the size of the tennis ball) via a drone. The trained model would be deployed in a Jetson Xavier NX device running inferences on the streaming video. I wonder if you have the training source code that I could reference as a starting point? I'd really appreciate it. Thanks.

[alessiocanepa]

Dear Wei, the network was trained using the object detection API of tensorflow via command line. We modified tensorflow's implementation of mobilenet in order to get the architecture presented in the paper, than we used directly the object detection command line API to train it. For this reason, unfortunately we still do not have a plug n play shareable package to retrain the network. Since the architecture of the network is simple, one idea is to implement it in keras starting from mobilenet and simply modifying its structure to get our network.

Thanks for the appreciation of the work, we hope it'll be helpful for your research.

Best,

Alessio

Il giorno gio 29 lug 2021 alle ore 16:11 Wei Teh @.***> ha scritto:

Hi, I came across your project on github after reading the paper: https://www.mdpi.com/1424-8220/21/4/1252/pdf and was very interested in diving deeper into this algorithm. I have a similar use case that involves tracking a relatively small object (about the size of the tennis ball) via a drone. The trained model would be deployed in a Jetson Xavier NX device running inferences on the streaming video. I wonder if you have the training source code that I could reference as a starting point? I'd be really appreciate it. Thanks.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/SEAlab-unige/TRexNet_demo/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACYW5WTXI6KJSTM6VQ2I2VTT2FOPPANCNFSM5BGRYXHA .

[sisrfeng]

Hi, would you mind sharing the entire training code now? Many thanks

[evekeen]

@alessiocanepa sharing the original code would significantly help further research!

[evekeen]

@alessiocanepa could you comment on how you connected your backbone to the SSD head? In SSD-Lite they connect C4, C5 layers and then add 4 extra layers also connected to SSD. From your paper, it looks like you only connect C5 layer and don't add any extra layers. Is that true?