🦎 VMT: Video Mask Transfiner 🦎

👉Novel highly efficient ViT structure for video instance segmentation.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅HD & more temporally stable mask
✅Higher resolution features for VIS
✅Detecting error-prone s-t. regions
✅Auto-refinement on training data!

More: https://bit.ly/3RKXtb4

🤯 #StableDiffusion + #Dallemini = BOOM! 🤯

👉A #colab notebook that combines Stable Diffusion + DALL-E Mini (Craiyon)

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

🐠VIS - Deformable Transformers 🐠

👉DeVIS: VIS method with efficiency and performance of deformable ViT

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Temp. multi-scale D-Attention
✅Instance-aware object queries
✅Mask: DA + multi-scale feats map
✅Improved multi-cue clip tracking
✅SOTA on YouTube-VIS 2021/OVIS

More: https://bit.ly/3TQv1Xc

🌈 X-NeRF: Cross-Spectral NeRF 🌈

👉Cross-Spectral NeRF from cams with different light spectrums

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅First ever cross-spectral NeRF
✅Avoiding non-trivial calib/match
✅Normalized Cross-Device Coords
✅Novel dataset w/ RGB, MS, & IR

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

👹TT-GNeRF: generative NeRF for Faces👹

👉TT-GNeRF: a novel 3D-aware GANs based on generative NeRF for faces

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅ETH + Uni_Trento + #Snap 🤯
✅DAEM for disentanglement of 3D model
✅"Training-as-Init, Optimizing-for-Tuning"
✅Consistency++, preserving non-target ROI
✅Unsupervised optimization of geometry

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

🎪 SOTA in Arbitrary Shape Text Detection 🎪

👉Novel unified coarse-to-fine Transformer for arbitrary shape text detection

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Coarse-to-fine arbitrary text detection
✅Accurate text detection, NO post-process
✅Boundary proposal generation mechanism
✅Innovative boundary transformer (iterative)
✅Boundary energy loss (BEL) for refinement

More: https://bit.ly/3D6Ryt4

🐲 Open-Source Self-Driving projects 🐲

👉A free repo with many autonomous vehicle-related projects

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Basic/Advance Lane/Line Detection
✅Driving behavior by training & validating
✅Autopilot: predicting steering angle

More: https://bit.ly/3qqJ7RB

🥤K-VIL: Keypoint-based visual imitation🥤

👉K-VIL: auto-incremental extraction of object-centric task representation.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Efficient task-relevant keypoints
✅Embodiment-independent tasks
✅Adaptation of tasks to new scenes
✅Input: only a small set of demo clips
✅Novel keypoint-based controller

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

💜 #Selfdriving in 80's. Damn Romantic 💜

👉The first self-driving car with people on board, 1986. So slow and lovely.

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

🏵️ TORAS: SOTA #AI for annotation 🏵️

👉TORAS: web-based AI-powered, cooperative, annotation platform.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅SOTA AI tools -> significant speedup
✅"Recipes" to define how to annotate
✅Repo with folder structure for storage
✅Also on-prem for (commercial) firms

More: https://bit.ly/3L78YI2

💮MAXIM: Multi-Axis MLP for Vision💮

👉#Google opens MAXIM, a multi-axis MLP for low-level vision

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Denoising, deblurring, dehazing, etc
✅Multi-axis gated MLP, linear complexity
✅Cross gating block, separate features
✅SOTA results on several datasets!

More: https://bit.ly/3Dmp8LI

🔥 A Survey on Diffusion Models 🔥

👉A comprehensive review of denoising diffusion models in #computervision 🤯

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Overview on diffusion models
✅Hot trend for the generative AI
✅A multi-perspective categorization
✅Current limitations / new directions

More: https://bit.ly/3RYG5zP

🔥🔥 UPDATE 🔥🔥

Code Released: https://github.com/sczhou/CodeFormer

🉐#AI finds where IG photos are taken🉐

👉Brilliant work of Depoorter, Belgium artist that handles #privacy, #AI & #socialmedia

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Recorded open cameras for weeks
✅Scraped all #Instagram photos
✅Matching Instagram vs. footage

More: https://bit.ly/3eL5dfc

🈯SAMURAI: in-the-wild Shape/Material🈯

👉#Google SAMURAI: shape, BRDF, per-image pose & illumination. Relightable #3D assets for #AR/#VR.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Parametrization for varying distances
✅Camera multiplex optimization
✅Posterior scaling of input images
✅Explicit meshes extraction with BRDF
✅Code/data soon available ->#NeurIPS

More: https://bit.ly/3BKWgf3

🟨 Lang<->Pics in 100+ Languages 🟨

👉#Google PaLI: unified lang-image #AI to perform tasks in 109 languages 🤯

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅PaLI: Pathways Lang & Image model
✅Answering, captioning, reasoning, etc
✅From Eng. to 109 lang. understanding
✅The new SOTA on several datasets

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

🍐PeRFception: Largest IR Dataset🍐

👉#Nvidia, a new frontier in data collection via Plenoxels: same info, -96.4% in size.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅POSTECH + NVIDIA + Caltech = 🤯
✅Size: -96.4% from original dataset!
✅2D/3D image/object class/semantic
✅Ready-to-use pipeline for implicit dataset

More: https://bit.ly/3eW9hJA

🐸 CHARL-E: Stable Diffusion in 1 click 🐸

👉CHARL-E packages Stable Diffusion into a simple app.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅No setup, dependencies, or internet
✅Images with 1-click on #macbook
✅Suitable only for M1/M2 processor
✅Source code under MIT license

More: https://bit.ly/3xv2z3G

🍋YOLOPv2: Better Driving Perception🍋

👉YOLOPv2: simultaneous object, road segmentation & lane detection

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅E2E perception net with better backbone
✅Efficient ELAN for reasonable memory
✅Stability for adapting to scenarios
✅SOTA on BDD100K, +50% faster!
✅Source code under MIT license

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

🍈SegNeXt: new SOTA in Semantic Seg.🍈

👉SOTA (by large margin) on ADE20K, Cityscapes, COCO-Stuff, Pascal VOC, Pascal Context, and iSAID 🤯

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Novel tailored network architecture
✅Spatial attention via multi-scale feats
✅Encoder + conv. better than transformers
✅SOTA on several datasets (ADE20K, etc.)

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