vikrant7 / mobile-vod-bottleneck-lstm

Implementation of Mobile Video Object Detection with Temporally-Aware Feature Maps using PyTorch
131 stars 37 forks source link

Usage of the Evaluate.py for the sequence data. #13

Open devarajnadiger opened 4 years ago

devarajnadiger commented 4 years ago

Hi,

I'm training the network using train_mvod_lstm1.py with a sequence of images(video) as an input(without any basenet or pretrained model) from the train_VID_seqs_list.txt. I got the models as well as checkpoints after completion of training. To predict on the test data set i'm using evaluate.py where it is reading individual frame and predicting for it.But i'm expecting that prediction should happen in sequence level that is read a set of frames as sequence and give a output. I have a question that will the evaluate.py in the existing repo. work for testing the data in sequence level or is it to predict individual frames only?

Thank you

faheuer commented 4 years ago

Push. I would like to understand the evaluate.py as well. When I test basenet after 30 epochs Imagenet VID I get diverged models

amiiiirrrr commented 4 years ago

Push. I would like to understand the evaluate.py as well. When I test basenet after 30 epochs Imagenet VID I get diverged models

hi faheuer how much your MAP? i trained mobilenetv2 ssd lite on imagenet VID and imagenet DET(30 common classes with VID) and MAP reached to 0.39.