🔥The power of Transformers🔥

👉100+ official implementations, papers, github repo and colab of:

01 GPT-Neo 2021
02 Transformer 2017
03 BERT 2018
04 GPT 2018
05 Univ.Transformer 2018
06 T-D 2018
07 GPT-2 2019
08 T5 2019
09 BART 2019
10 XLNet 2019
11...

👉The full list: https://github.com/ashishpatel26/Treasure-of-Transformers

⚕️Transformers in Medical ⚕️

👉100+ papers, implementations and code of Transformers in medical imaging.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Medical Image Segmentation
✅Medical Image Classification
✅Medical Image Reconstruction
✅Medical Image Registration
✅Medical Image Synthesis
✅Medical Image Detection
✅Clinical Report Generation
✅Survey and more..


The full list: https://bit.ly/3ILzswl

🪑#3D with Transformers🪑

👉ShapeFormer, transformer network for incomplete input

✅VQDIF representation for 3D
✅Transformer-based model
✅Partial input -> completed shape
✅ConvONet/Taming-transf/DCTransf.
✅SOTA for #3D shape completion

More: https://bit.ly/3s0D2f1

📱YOLO5 real-time logo detector📱

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Based on YOLOv5 family
✅Google Colab + Azure
✅90k pics, training: 2 weeks
✅Pretrained models/code
✅GNU License v3.0

More: https://bit.ly/3r8Qoa7

🦒RelTR: #AI scene-graphs🦒

👉One-stage method for object relationship via visual appearance only.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅RelTR ,end-to-end framework
✅Classifying dense relationships
✅Scene graphs on appearance only
✅No combining entities & labeling
✅Superior performance, faster

More: https://bit.ly/3r8k86Y

🥶SOTA in crowd analysis is INSANE🥶

👉Tencent unveils P2PNet to predict heads in images

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Pure point counting/detecting
✅Normalized Average Precision
✅VGG16-like architecture
✅Simultaneous point/confidence
✅License: only academic

More: https://bit.ly/33UjoK0

❄️OLSO: Transformers Optimization❄️

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Automagical with Hugging Face
✅GPU-based optimizations
✅Easily installation with pip
✅Apache License 2.0

More: https://bit.ly/3r8wY58

🦾SOTA in robotic manipulation🦾

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅VCD: Visible Connectivity Dynamics
✅VCG: Visible Connectivity Graph
✅Dynamics model over this VCG
✅Handling material, geometry, color
✅SOTA vs. model-based/model-free RL
✅Source code and models available

More: https://bit.ly/3HhusiH

📟VRT: new SOTA in super resolution📟


𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Image restoration via Swin
✅Residual Swin Transf. Blocks
✅SOTA in Artifact Reduction
✅SOTA in Super-resolution
✅SOTA in Denoising
✅Parameters -67%!
✅Non commercial 🥲

More: https://bit.ly/3rfAta1

🦖The new #MediaPipe is INSANE 🦖

👉Google just launched two new highly optimized body segmentation models

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Full body 3D pose
✅Designed for yoga, fitness & dance
✅Measurements for virtual tailor
✅Selfie Segmentation on call

More: https://bit.ly/3s6sjjx

🥸 Clothed avatars for #metaverse 🥸

👉Telepresence, AR/VR, anthropometry, and virtual try-on.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Differential loss of explicit mesh
✅Details via neural rendering
✅Explicit mesh updating
✅Consistency loss for quality++
✅Hi-Fi surfaces by S.S. optimization

More: https://bit.ly/3ohAN6d

🦕JoJoGAN: One Shot Face Stylization🦕

👉UIUC researchers unveil a novel method for one-shot image stylization.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Stylization from single input
✅Finetuning StyleGAN for stylization
✅No supervision, good generalization
✅MIT License (commercial allowed)

More: https://bit.ly/3ASVzyb

🧦SOTA in OOD detection for safer #AI🧦

👉Out-of-distribution (OOD) detection produces wrong/overconfident predictions.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Novel framework for OOD
✅Synthesizing virtual outliers
✅Novel unknown-aware training
✅Code and model available

More: https://bit.ly/3JnFIL9

🌅StyleGAN-XL neural synthesis🌅

👉From Tübingen, StyleGAN-XL: new SOTA for large diverse dataset.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅First 1024p-gen for large data
✅Growing strategy on StyleGAN3
✅Beyond the narrow domains
✅Pivotal Tuning Inversion (TPI)
✅SOTA vs. GAN & diffusion models

More: https://bit.ly/3HK9MQk

📌This keypoint is pure GLUE📌

👉Keypoints play a central role in computer vision.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Novel Object-centric keypoint
✅Novel sim2real training method
✅Intra-salience / inter-distinctness
✅Enforcing semantic consistency
✅Close to fully-supervised method!

More: https://bit.ly/3rth1qh

💡 LEDNet: seeing in the dark 💡

👉Researchers from NTU unveil LEDNet to see in the dark

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Novel data synthesis for low-light
✅Low-light/deblurring dataset
✅12k low-blur/normal-sharp pairs
✅LEDNet: lowlight + deblurring


More: https://bit.ly/3HIyYqM

👩‍🦰Back in the 50's with GAN👩‍🦰

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅A few thousand vintage faces
✅Models available for download
✅Stylegan2-ffhqu-1024x1024
✅NO Commercial allowed

More: https://bit.ly/3LlOyKX

🦠VNCA: bio-inspired generative model 🦠

👉A novel generative model loosely inspired by the biological processes of cellular growth and differentiation

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Variational Neural Cellular Automata
✅Probabilistic generative model
✅Learn from common vector format
✅Learn purely s.o. generative process
✅Far away from SOTA, but interesting

More: https://bit.ly/3oGb2wG

🍊Block-NeRF: Neural View Synthesis🍊

👉Large-scale scene reconstruction by multiple compact NeRFs that each fit into memory.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Berkeley + Google + Waymo = 🤯
✅Scaling NeRF to city-scale scenes
✅Trick: multiple simple NeRFs
✅Time decoupled, arbitrarily large scene
✅Data over months & different conditions

More: https://bit.ly/3GGVHBV

🥬HW-Accelerated Neuro-Evolution🥬

👉Scalable, general purpose, hardware accelerated neuro-evolution toolkit by Google

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Parallel on multiple TPU/GPUs
✅Neuro-evo algorithms with NNs
✅WaterWorld, Abstract paint, more
✅From Google, not an official product
✅Code under Apache License 2.0

More: https://bit.ly/3szEi9w