🗺️Neural Translation Image -> Map🗺️

👉A novel method for instantaneous mapping as a translation problem

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Bird’s-eye-view (BEV) map from image
✅A restricted data-efficient transformer
✅Monotonic attention from lang.domain
✅SOTA across several datasets

More: https://bit.ly/39MQ76Z

🥶 E2V-SDE: biggest troll ever? 🥶

👉E2V-SDE paper (accepted to #CVPR2022) consists of texts copied from 10+ previously published papers 😂

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Latent ODEs for Irregularly-Sampled TS
✅Stochastic Adversarial Video Prediction
✅Continuous Latent Process Flows
✅More papers....


More: https://bit.ly/3bsL8Zw (AUDIO ON!)

🔥🔥YOLOv6 is out: PURE FIRE!🔥🔥

👉YOLOv6 is a single-stage object detection framework for industrial applications

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Efficient Decoupled Head with SIoU Loss
✅Hardware-friendly for Backbone/Neck
✅520+ FPS on T4 + TensorRT FP16
✅Released under GNU General Public v3.0

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

🐪 BlazePose: Real-Time Human Tracking 🐪

👉Novel real-time #3D human landmarks from #google. Suitable for mobile.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅MoCap from single RGB on mobile
✅Avatar, Fitness, #Yoga & AR/VR
✅Full body pose from monocular
✅Novel 3D ground truth acquisition
✅Additional hand landmarks
✅Fully integrated in #MediaPipe

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

🔥YOLOv7: YOLO for segmentation🔥

👉YOLOv7: adding a lot of newer skills to the YOLO architecture family.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅YOLOv7, not a successor of YOLO family!
✅Framework for detection & segmentation
✅Applications based on #META detectron2
✅DETR & ViT detection out-of-box
✅Easy support for pipeline thought #ONNX
✅YOLOv4 + InstanceSegm. via single stage
✅The latest YOLOv6 training is supported!
✅Source code under GPL license.

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

🔥🔥 HD Dichotomous Segmentation 🔥🔥

👉 A new task to segment highly accurate objects from natural images.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅5,000+ HD images + accurate binary mask
✅IS-Net baseline in high-dim feature spaces
✅HCE: model vs. human interventions
✅Source code (should be) available soon

More: https://bit.ly/3ah2BDO

🔥🔥 Neural Segmentation on fire 🔥🔥

👉Novel methods for segmentation with mask calibration. Robustness++ in VOS.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Study: VOS robustness vs. perturbations
✅Adaptive object proxy (AOP) aggregation
✅Less errors due unstable pixel-level match
✅Code/models (should be) available soon

More: https://bit.ly/3yhIY6Q

😊😎 Seq-DeepFake via Transformers 😎😊

👉S-Lab opens Seq-DeepFake: Detecting Sequential DeepFake Manipulation

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Seq-DeepFake: sequences of facial edits
✅Dataset: 85k #deepfake manipulation
✅Powerful Seq-DeepFake Transformer
✅Code, dataset and models available!

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

🦒 Text2LIVE: Text-Driven Neural Editing 🦒

👉#Amazon unveils a novel #AI for text-driven edit of videos. Insane! 🤯

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Semantic edits of real-world videos
✅Edit layer–RGBA representing target
✅Edit layers synthesized on single input
✅No masks or a pre-trained generator

More: https://bit.ly/3NVP6aE

📟📟AI-Designed Circuits with Deep RL📟📟

👉#Nvidia unveils an #AI to design circuits from scratch, smaller and faster than SOTA ones

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Parallel prefix circuits for Hi-Perf
✅RL framework to explore the circuit space
✅Smaller, Faster, Power-- from the scratch

More: https://bit.ly/3yY9dk7

👽 Neural I2I with a few shoots 👽

👉#Alibaba unveils a novel portrait stylization. Limited samples (∼100) -> HD outputs

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Calibration first, translation later
✅Balanced distribution to calibrate bias
✅Spatially semantic constraints via geometry
✅Source code and models soon available!

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

🤹‍♂️ K-Means Mask Transformer 🤹‍♂️

👉#Google AI unveils kMaX-DeepLab, novel E2E method for segmentation

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅kMaX-DeepLab: k-means Mask Xformer
✅Rethinking relationship pixels / object
✅Cross-attention -> k-means clustering
✅The new SOTA on several dataset

More: https://bit.ly/3O2QV5I

☀️ 4D Neural Relightable Humans ☀️

👉Relighting4D: free-viewpoints relighting of humans under unknown illuminations

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Relight dynamic, free viewpoints
✅Disentangled reflectance/geometry
✅SOTA on synthetic/real datasets
✅Code/models under MIT License

More: https://bit.ly/3RF3yH9

🍰 Long-Term Object Segmentation 🍰

👉XMem: object segmentation for long clips with unified feature memory stores

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Inspired by Atkinson–Shiffrin model
✅Stores with different temporal scales
✅Memory consolidation algorithm
✅Compact/powerful long-term memory
✅Source code and models available

More: https://bit.ly/3PP0EOn

🔥🔥 Update 🔥🔥

👉Code https://github.com/THUDM/CogVideo

👉Demo https://wudao.aminer.cn/cogvideo/

More: https://bit.ly/3yP86BQ

🔥Grand Unification of Object Tracking🔥

👉UNICORN: unified method for SOT, MOT, VOS, & MOTS with a single neural net. 🤯

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Great unification for 4 tracking tasks
✅Bridging methods / pixel-wise corresp.
✅SOTA on 8 challenging benchmarks
✅Source code under MIT License

More: https://bit.ly/3o74h6g

🔥OmniBenchmark: CV beyond ImageNet🔥

👉 21 realms, 7,000+ concepts and 1M+ images. Far beyond ImageNet!

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅vs. ImageNet: 2.5x realms, 9x concepts
✅Conciseness: no concept overlapping
✅ReCo: Relational Contrastive Learning
✅New supervised contrastive learning SOTA

More: https://bit.ly/3RJRKU0

💣 HD Neural Avatar @130FPS 💣

👉Samsung unveils MegaPortraits: novel one-shot creation of HD neural human avatar

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅One-shot neural avatars, SOTA up 512p
✅"Upgrading" to megapixel via more pics
✅First Neural Head Avatars in HD
✅Up to to 130 FPS via #GPU

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

🔥Important update from #OpenAI🔥

👉 https://openai.com/blog/reducing-bias-and-improving-safety-in-dall-e-2/

🦚 TimeLens++: Event-based Interpolation 🦚

👉Novel event-based interpolation with non-linear flow & multi-scale fusion

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Novel motion spline estimator
✅Non-linear continuous event/frames flow
✅Multi-feature fusion, gated compression
✅Novel hybrid dataset with 100+ videos

More: https://bit.ly/3yJyY6g