1708.01241 - DSOD: Learning Deeply Supervised Object Detectors from Scratch.md

1708.01241 - DSOD: Learning Deeply Supervised Object Detectors from Scratch

• Other ref:

  • http://www.bijishequ.com/detail/535586
  • https://github.com/szq0214/DSOD
  • http://blog.csdn.net/jean_0724/article/details/77574403

• Introduction:

  • DSOD: learn object detectors from scratch
  • Combination of SSD and DenseNet

• Previous work focus on pretraining / fine tuning on ImageNet, cons:

  • Limited structure design space
  • Learning bias
  • Domain mismatch

• Our method:

  • Use backbone sub-network to extract features: modified DenseNet with a stem block, four dense blocks, two transition layers and two transition w/o pooling layers
  • Use front-end sub-network to predict over multi-scale response maps: fuses multi-scale prediction responses with an elaborated dense structure

• Design principle and result:

  • Proposal free: SSD
    • Faster RCNN and R-FCN can not converge, SSD can converge but worse than pretrained model
  • Deep supervision: DenseNet, alleviate gradient vanishing
    • Make pretrain-free more accurate than pretrained and fine-tuned SSD
  • Stem block: inspired by Inception-v3/v4
    • Better preformance
  • Dense prediction structure: SSD
    • A little slower than plain structure
    • More accuracy: increase by 0.4%
    • Less paramaters: decrease by 3.4M

• An strange observation: by using pretraining and fine tuning, the performance of DSOD is even lower(-0.4%) than the one without pretraining