Skip to content

Latest commit

 

History

History
75 lines (56 loc) · 2.28 KB

arch-variants.md

File metadata and controls

75 lines (56 loc) · 2.28 KB
Raw

Architecture Variants

In this work, we provide a set of variants to validate our design choices employed in DAVO. There are version names of these variants:

Version Description
v1-decay100k-sharedNN-dilatedPoseNN-cnv6_128-segmask_all-se_flow-abs_flow-fc_tanh DAVO
v1-decay100k-sharedNN-dilatedPoseNN-cnv6_128-no_segmask Ours w/o attention
v1-decay100k-sharedNN-dilatedPoseNN-cnv6_128-segmask_all-static Ours w/ static_attention
v1-decay100k-sharedNN-dilatedPoseNN-cnv6_128-no_segmask-se_insert Ours w/ feature attention
v1-decay100k-sharedNN-dilatedPoseNN-cnv6_128-segmask_all-se_seg_wo_tgt-fc_tanh DAVO (segmentation source)
v1-decay100k-sharedNN-dilatedPoseNN-cnv6_128-segmask_all-se_depth_wo_tgt_to_seg-fc_tanh DAVO (depth source)
v1-decay100k-sharedNN-dilatedPoseNN-cnv6_128-segmask_all-se_rgb_wo_tgt_to_seg-fc_tanh DAVO (rgb source)

you can quickly switch the variant by choosing the version name from them, or make your own variant by selecting components listed below while setting the command --version in the training phase and the testing phase.

Architecture Components

Choose the input of PoseNN:

  • v0 : only RGB frames.

  • v1 : the concatenation of RGB frames and optical flows.

    [additonally concatenate segmentation id maps]

    • -seglabelid

Choose learning rate decay:

  • -decay50k
  • -decay100k
  • -decay50k-staircase
  • -decay100k-staircase

Choose PoseNN:

  • -dilatedPoseNN

  • -dilatedCouplePoseNN

  • -sharedNN-dilatedPoseNN

  • -sharedNN-dilatedCouplePoseNN

    [adjust conv6 channels in PoseNN]

    • -cnv6_<num_channels>

    [adjust se_block between conv5 and conv6 in PoseNN]

    • -se_insert
    • -se_replace
    • -se_skipadd

Choose Attention mode:

  • -segmask_all
  • -segmask_rgb
  • -no_segmask

Choose AttentionNN source and layer types:

  • -se_flow

  • -se_spp_flow

  • -se_seg

  • -se_spp_seg

  • -se_mixSegFlow

  • -se_spp_mixSegFlow

  • -se_SegFlow_to_seg

  • -se_app_mixSegFlow

  • -se_seg_wo_tgt

  • -se_depth_wo_tgt_to_seg

  • -se_rgb_wo_tgt_to_seg

    [adjust the activated function in AttentionNN]

    • -fc_relu
    • -fc_tanh
    • -fc_lrelu