Skip to content

DataFusionBlock enhancements to extract and fuse optical flow features into the model - #17

Open
yusfcicek wants to merge 3 commits into
SysCV:mainfrom
yusfcicek:data_fusion_block_addition
Open

DataFusionBlock enhancements to extract and fuse optical flow features into the model#17
yusfcicek wants to merge 3 commits into
SysCV:mainfrom
yusfcicek:data_fusion_block_addition

Conversation

@yusfcicek

Copy link
Copy Markdown
  • In training and inference time extracts optical flow features in multi frames
  • Multiple frames are taken as a ring buffer, then per-frame optical flow features are extracted
  • Frames and optical flow features are fused by DataFusionBlock
  • Configuration improvements were made according to the number of datasets
  • Fixed bugs in torch functions in Python 3.9 and above

yusfcicek and others added 3 commits April 10, 2024 00:11
…lti frames

* Multiple frames are taken as a ring buffer, then per-frame optical flow features are extracted
* Frames and optical flow features are fused by DataFusionBlock
* Configuration improvements were made according to the number of datasets
* Fixed bugs in torch functions in Python 3.9 and above
Prepend a fork note to the README explaining what this branch adds on top
of SysCV/MaskFreeVIS, with upstream attribution kept intact.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015Tue2dg5QrW9Jwjcsqe94n
The two frames handed to calcOpticalFlowFarneback were converted to
grayscale with different colour orders -- COLOR_RGB2GRAY for the previous
frame and COLOR_BGR2GRAY for the next. The two conversions apply the
luminance weights to different channels, so consecutive frames landed in
different grayscale spaces and the estimated flow carried a systematic
error. Both now use COLOR_RGB2GRAY, matching the documented RGB input.

The same docstring named Lucas-Kanade while the code calls the Farneback
estimator, and the docstrings in build.py and base.py were copied from
detectron2's backbone equivalents and described building a backbone.

Also renames fusioned_* to fused_* and concanated_feature to
concatenated_feature, and fixes the wording in the OpticalFlowFusionBlock
docstring. All call sites are in this repository, so the keyword argument
rename is complete.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XZLNft11UJQfqcRXhcHnML
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant