Mobilenetv2 demo segmented image shift

david8862 / tf-keras-deeplabv3p-model-set

end-to-end DeepLab V3+ semantic segmentation pipeline, implemented with tf.keras

24 stars 8 forks source link

Mobilenetv2 demo segmented image shift #1

Open GotG opened 3 years ago

GotG commented 3 years ago

Hello

When attempting demo with mobilenetv2: python deeplab.py --model_type=mobilenetv2_lite --weights_path=/weights/deeplabv3_mobilenetv2_tf_dim_ordering_tf_kernels.h5 --classes_path=configs/voc_classes.txt --output_stride=16 --image

I get the following result. The segmentation is shifted.

result

Thanks!

david8862 commented 3 years ago

@GotG sorry for response delay. Which model file do you use, is it downloaded from the release link?

GotG commented 3 years ago

no probs. yes, the one from the link. deeplabv3_mobilenetv2_tf_dim_ordering_tf_kernels.h5

david8862 commented 3 years ago

Is it from https://github.com/david8862/tf-keras-deeplabv3p-model-set/releases/download/1.0.0/deeplabv3p_mobilenetv2_lite_512_os16_voc.tar.gz? If true then looks quite weired. Could you provide the original test image? I can try to verify on my testbed