1.1.0 Added new models and datasets, updated preprocessing and more

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)

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

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