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.
-
v0 : only RGB frames.
-
v1 : the concatenation of RGB frames and optical flows.
- -seglabelid
- -decay50k
- -decay100k
- -decay50k-staircase
- -decay100k-staircase
-
-dilatedPoseNN
-
-dilatedCouplePoseNN
-
-sharedNN-dilatedPoseNN
-
-sharedNN-dilatedCouplePoseNN
- -cnv6_<num_channels>
- -se_insert
- -se_replace
- -se_skipadd
- -segmask_all
- -segmask_rgb
- -no_segmask
-
-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
- -fc_relu
- -fc_tanh
- -fc_lrelu