furkanom/tf-ssd

Tensorflow 2 single shot multibox detector (SSD) implementation from scratch with MobileNetV2 and VGG16 backbones

tf-ssd's Issues

problem about the bbox format

I need to feed my own data and I feed the data as the note under 'calculate_actual_outputs':

And my data is like this:
the prior boxes:

the ground truth boxes:

I think my data format is correct, but I got this error:

It seems that the 'pos_cond' should be shape [batch, num_of_anchers, 4], is this caused by my data format?

TRAINING ON A CUSTOM DATASET

i am wondering what changes are needed to train the model on a custom dataset.
thank you

include_top=False seems not work

No matter I change the setting of ‘include_top’ to True or False, the base net always has the last Dense layer.
Anyone got something about this?

Provide pretrained model

Thank for your project, it is very useful for me. I want to test inference with this model first, It will be greate if you provide pretrained tf.keras h5 model.

Support COCO dataset?

Great work! Do you plan to support mscoco dataset?