🐦 EfficientVIS: new SOTA for VIS 🐦

👉Simultaneous classification, segmentation, and tracking multiple object instances in videos

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Efficient and fully end-to-end
✅Iterative query-video interaction
✅First RoI-wise clip-level RT-VIS
✅Requires 15× fewer epochs

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

🐠#AI-clips from single frame🐠

👉Moving objects in #3D while generating a video by a sequence of desired actions

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅A playable environments
✅A single starting image🤯
✅Controllable camera
✅Unsupervised learning

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

🧊Kubric: AI dataset generator🧊

👉Open-source #Python framework for photo-realistic scenes: full control, rich annotations, TBs of fresh data 🤯

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Synthetic datasets with GT
✅From NeRF to optical flow
✅Full control over data
✅Ok privacy & licensing
✅Apache License 2.0

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

🪂µTransfer for enormous NNs 🪂

👉Microsoft unveils how to tune enormous neural networks

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅New HP tuning: µTransfer
✅Zero-shot transfer to full-model
✅Outperforming BERT-large
✅Outperforming 6.7B GPT-3
✅Code under MIT license

More: https://bit.ly/3qc37Ij

🐧Semantic via only text supervision🐧

👉GroupViT with a text encoder on a large-scale image-text dataset: semantic with any pixel-level annotations in training!

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Hierarc. Grouping Vision Transf.
✅Additional text encoder
✅NO pixel-level annotations
✅Semantic-seg task via zero-shot
✅Source code available soon

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

⌚4D-Net: Lidar + RGB synchronization⌚

👉Google unveils 4D-Net to combine 3D LiDAR and onboard RGB camera

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Point clouds/images in time
✅Fusing multiple modalities in 4D
✅Novel sampling for 3D P.C. in time
✅New SOTA for 3D detection

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

🐌 New SOTA in video synthesis! 🐌

👉Snap unveils a novel multimodal video generation framework via text/images

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Multimodal video generation
✅Bidirectional transformer
✅Video token with self-learn.
✅Text augmentation for robustness
✅Longer sequence synthesis

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

🎁 StyelNeRF source code is out 🎁

👉3D consistent photo-realistic image synthesis

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅NeRF + style generator
✅3D consistency for HD image
✅Novel regularization loss
✅Camera control on styles

More: https://bit.ly/3t5xC49

🦎CLD-based generative #AI by #Nvidia🦎

👉Nvidia unveils a novel critically-damped Langevin diffusion (CLD) for synthetic data

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅A novel diffusion process for SGMs
✅Novel score matching obj. for CLD
✅Hybrid denoising score matching
✅Efficient sampling from CLD model
✅Source code under a specific license

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

🛸UFO: segmentation @140+ FPS🛸

👉Unified Transformer Framework for Co-Segmentation, Co-Saliency & Salient Object Detection. All in one!

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Unified framework for co-segmentation
✅Co-segmentation, co-saliency, saliency
✅Block for long-range dependencies
✅Able to reach for 140 FPS in inference
✅The new SOTA on multiple datasets
✅Source code under MIT License

More: https://bit.ly/3KLd9b9

👗 Multi-GANs fashion 👗

👉Global GAN blended with other GANs for faces, shoes, etc.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Multi-GAN framework
✅Several generators
✅Free of artifacts
✅Full-body generation
✅Humans, 1024x1024

More: https://bit.ly/37mfOte

🚧 FLAG: #3D Avatar Generation 🚧

👉A flow-based generative model of the 3D human body from sparse observations.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅FLow-based Avatar Generative
✅Conditional distro of body pose
✅Exact pose likelihood process
✅Invertibility -> oracle latent code

More: https://bit.ly/3CQpk3p

💃 Dancing in the wild with StyleGAN 💃

👉StyleGAN-based animations for AR/VR apps

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Video based motion retargeting
✅A StyleGAN architecture based
✅Novel explicit motion representation
✅SOTA qualitatively & quantitatively

More: https://bit.ly/3CZbL1W

🪀TensoRF: the 4D evolution of NeRF 🪀

👉TensoRF, a novel radiance fields via 4D-tensor: 3D voxel grid with per-voxel multi-channel feats.

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅VM decomposition technique
✅Low-rank tensor factorization
✅Lower memory footprint (speed)
✅TensoRF is the new SOTA in R.F.
✅Code under the MIT License

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

🔼 GAN-meshes without key-points 🔼

👉ETH unveils a GAN framework for generating textured triangle meshes without annotations

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Generative of textured meshes
✅3D generator for all categories
✅3D pose estimation framework
✅Code licensed under MIT License

More: https://bit.ly/3qfH9nJ

🐯 S.S. Latent Image Animator 🐯

👉Self-supervised autoencoder to animate unseen images by linear navigation in latent

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Latent Image Animator
✅Linear displacement in latent
✅SOTA: VoxCeleb, Taichi, TED-talk
✅Source code (soon) available

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

🪨 Google URF for neural-synthesis 🪨

👉Sequence of RGB + Lidar -> 3D surfaces and novel RGB images synthesized

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Extending Neural Radiance Fields
✅Leveraging asynch. lidar data
✅Addressing exposure variation
✅Leveraging segmentations for sky
✅SOTA #3D reconstructions/synthesizes

More: https://bit.ly/3L2vTDb

🚛 AV2: next-gen. self driving 🚛

👉One of the biggest dataset ever for #autonomousdriving

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅1k seq. of multimodal data
✅3D annotations, 26 categories
✅20k lidar & map-aligned pose
✅250k challenging interactions
✅HD Map: 3D lane & crosswalk
✅CC BY-NC-SA 4.0 license

More: https://bit.ly/3trx3lw

🤖CaTGrasp in Clutter from Simulation🤖

👉Task-relevant grasping: trained solely in simulation with synthetic + SS. hand-object interaction

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Novel cat-level, relevant grasping
✅S.S. hand-object-contact
✅Tiny objects from dense clutter
✅Train-simulation -> to real
✅Source code under Apache 2.0

More: https://bit.ly/3L2YVCo

🛼 Drive & Segment without Supervision 🛼

👉Learning pixel-wise semantic seg. on non-curated data collection by cars (cameras + LiDAR) driving around a city

𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬:
✅Cross-modal unsupervised
✅Synchronized LiDAR & RGB
✅Object proposal on LiDAR points
✅SOTA, significant improvements

More: https://bit.ly/3L0wWTW