computer vision deep learning weight transfer

Learning to Segment Every Thing

weight transfer

Posted by stephen_zhou on December 9, 2017

result.png

mask rcnn

mask_rcnn.png

architecture

architecture.png

training

Stage-wise training: First stage: train a Faster R-CNN Second stage: train mask head and weight transfer with box head fixed

End-to-end joint training: Jointly train the bounding box head and the mask head But stop the gradient with the respect to $ W^c_{det} $

Ablation Experiment

Train on COCO by partitioning the 80 classes into sets A and B, 20/60 split , 20 contained in PASCAL VOC and the 60 that are not

image.png

image.png

image.png

image.png

Large-Scale Instance Segmentation

A from the COCO dataset, B from the Visual Genome (VG) dataset VG: 108077 images , over 7000 category synsets annotated with object bounding boxes image.png

Conclusion:

  • partially supervised
  • Weight transfer
  • Multi-task
CATALOG
    FEATURED TAGS
    machine learning generative model computer vision deep learning GAN VAE object detection Video Understanding Generative Model RNN Neural Networks Neural Network Image Synthesis
    FRIENDS