TorchVision Roadmap - 2021 H2

[NicolasHug]

This issue is be used to keep track of the roadmap of TorchVision for the second half of 2021 and get feedback from our users.

For the 2021 H1 roadmap, please see #3221

• Batteries included:
  • Scope is at [RFC] TorchVision with Batteries included - Phase 1 #3911
• Datasets:
  • New datasets design supporting DataPipes
• Video
  • Video GPU Decoding ([RFC] Hardware-accelerated video decoding #2439)
  • Optical Flow support: RAFT implementation with datasets and training /evaluation reference and pre-trained weights RAFT model and training reference #4644
  • New Video datasets e.g. Charades / SSV2 / EpicKitchen (pending discussion)
• IO:
  • Faster image decoding. Possible tracks: A100 support (Potential improvements to jpeg decoding on GPU #3848), Batch-decoding support, better libjpeg-turbo integration
• Perf:
  • Support for SIMD instructions for ops
• Docs:
  • More gallery examples (finish porting the examples folder)
  • Possibly move the vision tutorials in the torchvision gallery
• Items moved from 2021 H1:
  • Rotated Boxes ([RFC] Rotated Bounding Boxes #2761)

[shakeebmurtaza]

@NicolasHug I can add transformer models in torch.vision as indicated above.

[oke-aditya]

Guess we should now look at 2022 H1 ? 😃 cc @datumbox @NicolasHug

[datumbox]

@oke-aditya Thanks for the ping. Yes, we need to finalize and draft a roadmap proposal for 2022H1. We'll do it soon.