Skip to content

Releases: MichiganCOG/ViP

1.1.0 Added new models and datasets, updated preprocessing and more

15 Nov 18:12
74776f2
Compare
Choose a tag to compare

New features:

  • Supports both box (4 coordinate) and key-point (2 coordinate) annotations
  • Updated extract clips preprocessing options
  • Added video preprocessing for scaling/zoom and translation
  • Added gradient clipping
  • Resume training

Added Models:

  • I3D (activity recognition)
  • DVSA (+fw, obj) (video object grounding)

Added Datasets:

  • YouCook2-Bounding Boxes (video object grounding)
  • DHF1K (video saliency prediction)
  • KTH (activity recognition)
  • Manual Hands (hand key-point detection)

Bug fixes:

  • Clip stride must be greater than 0
  • Reset validation accuracy between epochs
  • Matching training and validation batch sizes for plots
  • Update COCO json generation to python3

1.0.1: Update json template frame size, fix num_clip=0 bug (#3)

13 Aug 20:23
Compare
Choose a tag to compare

Bugfix
Major bug fix when num_clips = 0 to generate multiple clips of a given clip length

Addition
Link to Wiki along with basic FAQ. Will continue to add more to the wiki page

Video Platform: A Pytorch-based video handler with a common user-interface for ANY problem domain

12 Aug 15:04
0326d61
Compare
Choose a tag to compare

Version 1.0

  • Training and Evaluation files are completed
  • Data loading pipeline
  • Config and argument reader
  • Checkpoint saving and loading
  • Implemented datasets: HMDB51, ImageNetVID, MSCOCO, VOC2007
  • Implemented recognition models: C3D
  • Implemented detection models: SSD
  • Implemented metrics: Recognition accuracy, IOU, mAP, AP
  • Implemented losses: MSE, Cross entropy

Version 0.1

  • Basic training file for recognition is ready, next step is to work on data loader since that is the immediate function being loaded
  • Began inclusion of dataset loading classes into datasets folder