🚛 DeepETA: #Uber ETA via #AI🚛

👉Uber unveils the low-latency deep architecture for global ETA prediction

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Latency / Accuracy / Generality
✅7 NNs architectures tested
✅Encoder-decoder + Self-Attention
✅Linear transformer (kernel trick)
✅Feature sparsity for speed

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

✏️CLIPasso: Semantic Sketching via CLIP✏️

👉Sketching method guided by geometric and semantic simplifications (CLIP)

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅EPFL, TAU and IDC Herzliya
✅CLIP image encoder for sketching
✅Sketching as a set of Bezier curves
✅Param-optimization on CLIP-loss
✅Source code and models available

More: https://bit.ly/3oLEDF4

🪂SAHI: slicing detection/segmentation🪂

👉An open-source lightweight library for large scale object detection & instance segmentation

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Slicing Aided Hyper Inference
✅Large-scale detection/segment.
✅Sliced inference and merging
✅Utils for conversion, slicing, etc.
✅Code licensed under MIT License

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

🎁100,000,000 image-text pairs!🎁

👉Large-scale Chinese cross-modal dataset for benchmarking different multi-modal pre-training methods.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅100 Million <image, text> pairs
✅>200px size, aspect ratio (1/3~3)
✅Models of ResNet, ViT & SwinT
✅Methods: CLIP, FILIP and LiT
✅Privacy/Sensitive words 🤔

More: https://bit.ly/34BqlzX

🧁33 Million synthetic pedestrians🧁

👉A novel large, fully synthetic dataset

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Exploiting the #gta5 engine
✅764 full-HD videos @20 fps
✅33M+ person instances
✅BBs & segmentation masks
✅2D/3D keypoints & depth

More: https://bit.ly/36njlY1

🥝Marker-free 6D-point tracking🥝

👉Full position and rotation of skeletal joints, with only a RGB frame

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Full 3-axis joint rotations
✅V-markers, emulating mocap
✅#3D from monocular with NN
✅Generalization, no retraining
✅SOTA rotation/position est.

More: https://bit.ly/34GdoF5

🧼 Synthetic dataset for #Retail 🧼

👉A large-scale photorealistic synthetic dataset with annotations for semantic segmentation, instance segmentation, depth estimation, and object detection.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Dataset from Standard.AI
✅2,134 unique scenes
✅25k+ annotated samples
✅Introducing the "change detection"
✅Multi-view representation learning
✅NonCommercial-ShareAlike 4.0

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

🌈 Graph Neural Nets Forecasting🌈

👉Data-driven approach for forecasting global weather using graph neural networks

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Data-driven forecasting via GNNs
✅Model: 6.7M parameters, float32
✅6-hours forecast in 0.04 secs.
✅A 5-day forecast in 0.8 secs.

More: https://bit.ly/3LH4CXR

🥫Watch Those Words!🥫

👉Berkeley unveils a novel approach to discover cheap-fake and visually persuasive deep-fakes

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Regardless of falsification
✅Semantic person-specific
✅Word-conditioned analysis
✅Generalization across fakes

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

🔋V2X-sim for #selfdriving is out!🔋

👉V2X: collaboration between a vehicle and any surrounding entity

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Suitable for #selfdrivingcars
✅Rec. from road & vehicles
✅Multi-streams/perception
✅Detection, tracking, & segmentation
✅RGB, depth, semantic, BEV & LiDAR

More: https://bit.ly/3H6veOI

🍏Infinite Synthetic dataset for Fitness🍏

👉Opensource synthetic images for fitness, single/multi-person, and realistic variation in lighting, camera angles, and occlusions

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅60k images, 1-5 avatars
✅15 categories, 21 variations
✅Blender and ray-tracing
✅SMPL-X + facial expression
✅Cloth/skin tone sampled
✅147 4K HDRI panoramas
✅Creative Commons 4.0

More: https://bit.ly/33B1R9q

♊ DITTO: Digital Twins from Interaction ♊

👉Digitizing objects for #metaverse through interactive perception

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅DIgital Twin of arTiculated Objects
✅Geometry & kinematic articulation
✅Articulation & 3D via perception
✅Source code under MIT License

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

🤖 Robotic Telekinesis from Youtube 🤖

👉CMU unveils a Robot that observes humans and imitates their actions in real-time

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Enabling robo-hand teleoperation
✅Suitable for untrained operator
✅Single uncalibrated RGB camera
✅Leveraging unlabeled #youtube
✅No active fine-tuning or setup
✅No collision via Adv-Training

More: https://bit.ly/3H7zUnh

💄DIGAN: #AI for video generation💄

👉A novel INR-based generative adversarial network for video generation

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Dynamics-aware generator
✅INR-based clip generator
✅Manipulating space/time
✅Identifying unnatural motion

More: https://bit.ly/3H6sHE4

🦄FILM Neural Frame Interpolation🦄

👉Frame interpolation that synthesizes multiple intermediate frames from two input images with large in-between motion

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Single unified network
✅High quality output
✅SOTA on the Xiph
✅Apache License 2.0

More: https://bit.ly/3pl4ZxH

🔈Neural Maintenance via listening🔈

👉Novel neural-method to detect whether a machine is "healthy" or requires maintenance

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Defects at an early stage
✅FDWT, fast discrete wavelet
✅Learnable wavelet/denoising
✅Unsupervised learnable FDWT
✅The new SOTA in PM

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

🟦🟨 StyleGAN on Internet pics 🟦🟨

👉StyleGAN on raw uncurated images collected from Internet

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Outliers & multi-modal
✅Self-distillation approach
✅Self-filtering of outliers
✅Perceptual clustering

More: https://bit.ly/33Z1d5H

🦜The new SOTA for Unsupervised 🦜

👉Self-supervised transformer to discover objects in images

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Visual tokens as nodes in graph
✅Edges as connectivity score
✅The second smallest eV = fg
✅Suitable for unsupervised saliency
✅Weakly supervised obj. detection
✅Code under MIT License


More: https://bit.ly/3sqbFg3

🥦 GAN-generated CryptoPunks 🥦

👉A simple (and funny) SN-GAN to generate cryptopunks

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Spectral normalization (2018)
✅Easy to incorporate into training
✅A project by Teddy Koker 🎩

More: https://bit.ly/35C1rQI

🤪SEER: self-AI from BILLIONS pic🤪

👉META + INRIA trained models on billions of random images without any pre-processing or assumptions

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Self-supervised on pics from web
✅Discovering properties in datasets
✅More fair, less biased & less harmful
✅Better OOD generalization
✅Source code available!

More: https://bit.ly/3vy69dd