Generate TF-record and training with Deeplab v3 +

[tuanng...@gmail.com]

Hi everyone, 

Im training deeplab with lvis dataset and have an issue :

I generate tfrecord file with only 3 class : person, dog,cat. Using pretrained model from imageNet. After training ( about 60000 steps) , the model only segments 1 class in a picture. Example picture have person and dog, it only regconizes dog and color of person is black. I check the annotation file of lvis and realize not all of picture which have person have segmentation-annotation.

Best regard,

Tobi

[Ross Girshick]

I check the annotation file of lvis and realize not all of picture which have person have segmentation-annotation.

Annotations are necessarily sparse and the annotation completeness guarantee is at the (image, category) pair level. Please read the paper for more information: https://arxiv.org/abs/1908.03195.

[dgks0n]

Hi Ross,

Bases on segmentation annotations, do you think that will make some noise during the training model? For example an image was included a person but the corresponding ground-truth did not provide segmentation for the whole person, just shoes. So, the model will confuse due to missing annotations.