Icevision: Video Object Segmentation Support

[rsomani95]

Stemming from this discussion on the Discord chat, we should eventually support video data loading and models in icevision.

Models

Here are two state of the art papers in the space:

1. CFBI [GitHub] [Paper]
2. STM (Space-Time Memory) Network [GitHub] [Paper]

PapersWithCode for Video Object Segmentation (VOS)

For CFBI, the training code is not released, but the author mentions that we could email them asking for the same. The repo has a BSD license, so I presume they'd be open to integrating a training pipeline with icevision?

Datasets

As for datasets, there's two popular datasets models are typically benchmarked against in this space:

1. DAVIS
2. YouTube-VOS

I'm somewhat familiar with the data format for YouTube-VOS -- they provide every 5th frame of the video as a JPEG image, so there's no real video loading required there. However, we should eventually build an base API that can load frames from videos and adapters for different datasets?

---

VOS models typically take an input mask that they track for the duration of the video. Eventually, it would be awesome to pipe the outputs from a Mask-RCNN like model into a VOS model.

[rsomani95]

Relevant library: https://github.com/visionml/pytracking