This media is not supported in your browser
VIEW IN TELEGRAM
๐#3D scene manipulation from 2D๐
๐Reconstruct, decompose, manipulate & render 3D scenes in a single pipeline
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Unique 3D, non-occupied space from 2D
โ Inverse query algorithm for shapes
โ First synthetic dataset for 3D editing
More: https://bit.ly/3RlYhTY
๐Reconstruct, decompose, manipulate & render 3D scenes in a single pipeline
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Unique 3D, non-occupied space from 2D
โ Inverse query algorithm for shapes
โ First synthetic dataset for 3D editing
More: https://bit.ly/3RlYhTY
๐ฅ11โค1
This media is not supported in your browser
VIEW IN TELEGRAM
๐StableFace: Talking Face Generation๐
๐Analysis on motion jittering in 3D face generation (audio-in -> video-out)
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Motion jittering analysis for stability
โ Gaussian-based adaptive smoothing
โ Augmented erosions of neural renderer
โ Audio-fused generator for dependency
More: https://bit.ly/3Kt95gI
๐Analysis on motion jittering in 3D face generation (audio-in -> video-out)
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Motion jittering analysis for stability
โ Gaussian-based adaptive smoothing
โ Augmented erosions of neural renderer
โ Audio-fused generator for dependency
More: https://bit.ly/3Kt95gI
๐5๐ฑ3โค1
This media is not supported in your browser
VIEW IN TELEGRAM
๐งก Avatarization in 90's. So Romantic ๐งก
๐Making of the first #MortalKombat in early 90's
More: https://bit.ly/3wTSpJB
๐Making of the first #MortalKombat in early 90's
More: https://bit.ly/3wTSpJB
โค13
This media is not supported in your browser
VIEW IN TELEGRAM
๐ Massive Dataset in Virtual Cities ๐
๐Synthehicle: 7 hours of labeled material, 340 cams, 64 days, rain, dawn, & night scenes.
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Multi-target multi-cam tracking
โ 2D, 3D, segm. & depth annotations
โ Instance, semantic & panoptic segm.
โ 340 clips, 64 scenes, 17 hrs, 4M BBs
More: https://bit.ly/3TArHiV
๐Synthehicle: 7 hours of labeled material, 340 cams, 64 days, rain, dawn, & night scenes.
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Multi-target multi-cam tracking
โ 2D, 3D, segm. & depth annotations
โ Instance, semantic & panoptic segm.
โ 340 clips, 64 scenes, 17 hrs, 4M BBs
More: https://bit.ly/3TArHiV
โค10๐6
This media is not supported in your browser
VIEW IN TELEGRAM
๐ชจControllable #3D Adversarial Face๐ชจ
๐#Meta (+CMU) on decoupling identity/expression + granular control over expressions
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Supervised auto-enc. + GAN
โ UV texture maps + 3D faces
โ Control expression, saving ID
โ Code under X11 License
More: https://bit.ly/3AVE80q
๐#Meta (+CMU) on decoupling identity/expression + granular control over expressions
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Supervised auto-enc. + GAN
โ UV texture maps + 3D faces
โ Control expression, saving ID
โ Code under X11 License
More: https://bit.ly/3AVE80q
๐6
This media is not supported in your browser
VIEW IN TELEGRAM
๐ฅ DALLยทE: Outpainting via #NLP ๐ฅ
๐Extending any original image, creating large-scale images in any aspect ratio
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Extending an image beyond its borders
โ Visual elements in same style of the input
โ Driving the image "story" in new directions
โ Shadows, reflections & textures w/ context
More: https://bit.ly/3eoH8uD
๐Extending any original image, creating large-scale images in any aspect ratio
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Extending an image beyond its borders
โ Visual elements in same style of the input
โ Driving the image "story" in new directions
โ Shadows, reflections & textures w/ context
More: https://bit.ly/3eoH8uD
๐ฅ20๐คฏ7โค1
This media is not supported in your browser
VIEW IN TELEGRAM
๐ช๏ธ TimeLapse++: Video Temporal Pyramid๐ช๏ธ
๐Multi-scale lens to view the passage of time: far beyond a "classic" timelapse
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Inspired by "old-school" spatial pyramids
โ Video Spectrogram to go through pyramid
โ Months/years of data in a few seconds!
โ Multi-temporal freq., no aliasing
More: https://bit.ly/3TKnYPS
๐Multi-scale lens to view the passage of time: far beyond a "classic" timelapse
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Inspired by "old-school" spatial pyramids
โ Video Spectrogram to go through pyramid
โ Months/years of data in a few seconds!
โ Multi-temporal freq., no aliasing
More: https://bit.ly/3TKnYPS
๐คฏ6๐2โค1
This media is not supported in your browser
VIEW IN TELEGRAM
๐ซ Stable Diffusion Video is out! ๐ซ
๐A free notebook to generate videos by interpolating the latent space of SD.
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Blueberry to strawberry spaghetti
โ Dream items from same prompt
โ Morph different prompts (seeds)
โ Built on a script by A. Karpathy
More: https://bit.ly/3ey8632
๐A free notebook to generate videos by interpolating the latent space of SD.
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Blueberry to strawberry spaghetti
โ Dream items from same prompt
โ Morph different prompts (seeds)
โ Built on a script by A. Karpathy
More: https://bit.ly/3ey8632
๐คฏ15๐1
This media is not supported in your browser
VIEW IN TELEGRAM
๐ฆ VMT: Video Mask Transfiner ๐ฆ
๐Novel highly efficient ViT structure for video instance segmentation.
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ HD & more temporally stable mask
โ Higher resolution features for VIS
โ Detecting error-prone s-t. regions
โ Auto-refinement on training data!
More: https://bit.ly/3RKXtb4
๐Novel highly efficient ViT structure for video instance segmentation.
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ HD & more temporally stable mask
โ Higher resolution features for VIS
โ Detecting error-prone s-t. regions
โ Auto-refinement on training data!
More: https://bit.ly/3RKXtb4
๐คฏ9โค1
๐คฏ #StableDiffusion + #Dallemini = BOOM! ๐คฏ
๐A #colab notebook that combines Stable Diffusion + DALL-E Mini (Craiyon)
More: https://bit.ly/3TTOshR
๐A #colab notebook that combines Stable Diffusion + DALL-E Mini (Craiyon)
More: https://bit.ly/3TTOshR
๐ฅ9๐5๐ข1
This media is not supported in your browser
VIEW IN TELEGRAM
๐ VIS - Deformable Transformers ๐
๐DeVIS: VIS method with efficiency and performance of deformable ViT
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Temp. multi-scale D-Attention
โ Instance-aware object queries
โ Mask: DA + multi-scale feats map
โ Improved multi-cue clip tracking
โ SOTA on YouTube-VIS 2021/OVIS
More: https://bit.ly/3TQv1Xc
๐DeVIS: VIS method with efficiency and performance of deformable ViT
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Temp. multi-scale D-Attention
โ Instance-aware object queries
โ Mask: DA + multi-scale feats map
โ Improved multi-cue clip tracking
โ SOTA on YouTube-VIS 2021/OVIS
More: https://bit.ly/3TQv1Xc
๐ฅ8โค1๐1
This media is not supported in your browser
VIEW IN TELEGRAM
๐ X-NeRF: Cross-Spectral NeRF ๐
๐Cross-Spectral NeRF from cams with different light spectrums
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ First ever cross-spectral NeRF
โ Avoiding non-trivial calib/match
โ Normalized Cross-Device Coords
โ Novel dataset w/ RGB, MS, & IR
More: https://bit.ly/3RqHnUo
๐Cross-Spectral NeRF from cams with different light spectrums
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ First ever cross-spectral NeRF
โ Avoiding non-trivial calib/match
โ Normalized Cross-Device Coords
โ Novel dataset w/ RGB, MS, & IR
More: https://bit.ly/3RqHnUo
๐7
This media is not supported in your browser
VIEW IN TELEGRAM
๐นTT-GNeRF: generative NeRF for Faces๐น
๐TT-GNeRF: a novel 3D-aware GANs based on generative NeRF for faces
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ ETH + Uni_Trento + #Snap ๐คฏ
โ DAEM for disentanglement of 3D model
โ "Training-as-Init, Optimizing-for-Tuning"
โ Consistency++, preserving non-target ROI
โ Unsupervised optimization of geometry
More: https://bit.ly/3ARZmMw
๐TT-GNeRF: a novel 3D-aware GANs based on generative NeRF for faces
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ ETH + Uni_Trento + #Snap ๐คฏ
โ DAEM for disentanglement of 3D model
โ "Training-as-Init, Optimizing-for-Tuning"
โ Consistency++, preserving non-target ROI
โ Unsupervised optimization of geometry
More: https://bit.ly/3ARZmMw
๐ฅ4โค1๐1
๐ช SOTA in Arbitrary Shape Text Detection ๐ช
๐Novel unified coarse-to-fine Transformer for arbitrary shape text detection
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Coarse-to-fine arbitrary text detection
โ Accurate text detection, NO post-process
โ Boundary proposal generation mechanism
โ Innovative boundary transformer (iterative)
โ Boundary energy loss (BEL) for refinement
More: https://bit.ly/3D6Ryt4
๐Novel unified coarse-to-fine Transformer for arbitrary shape text detection
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Coarse-to-fine arbitrary text detection
โ Accurate text detection, NO post-process
โ Boundary proposal generation mechanism
โ Innovative boundary transformer (iterative)
โ Boundary energy loss (BEL) for refinement
More: https://bit.ly/3D6Ryt4
โค8๐2๐ข1
This media is not supported in your browser
VIEW IN TELEGRAM
๐ฒ Open-Source Self-Driving projects ๐ฒ
๐A free repo with many autonomous vehicle-related projects
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Basic/Advance Lane/Line Detection
โ Driving behavior by training & validating
โ Autopilot: predicting steering angle
More: https://bit.ly/3qqJ7RB
๐A free repo with many autonomous vehicle-related projects
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Basic/Advance Lane/Line Detection
โ Driving behavior by training & validating
โ Autopilot: predicting steering angle
More: https://bit.ly/3qqJ7RB
๐ฅ22๐1
This media is not supported in your browser
VIEW IN TELEGRAM
๐ฅคK-VIL: Keypoint-based visual imitation๐ฅค
๐K-VIL: auto-incremental extraction of object-centric task representation.
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Efficient task-relevant keypoints
โ Embodiment-independent tasks
โ Adaptation of tasks to new scenes
โ Input: only a small set of demo clips
โ Novel keypoint-based controller
More: https://bit.ly/3eIrxpP
๐K-VIL: auto-incremental extraction of object-centric task representation.
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Efficient task-relevant keypoints
โ Embodiment-independent tasks
โ Adaptation of tasks to new scenes
โ Input: only a small set of demo clips
โ Novel keypoint-based controller
More: https://bit.ly/3eIrxpP
๐ฅ7๐1
This media is not supported in your browser
VIEW IN TELEGRAM
๐ #Selfdriving in 80's. Damn Romantic ๐
๐The first self-driving car with people on board, 1986. So slow and lovely.
More: https://bit.ly/3BtRDon
๐The first self-driving car with people on board, 1986. So slow and lovely.
More: https://bit.ly/3BtRDon
โค9๐4๐3
This media is not supported in your browser
VIEW IN TELEGRAM
๐ต๏ธ TORAS: SOTA #AI for annotation ๐ต๏ธ
๐TORAS: web-based AI-powered, cooperative, annotation platform.
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ SOTA AI tools -> significant speedup
โ "Recipes" to define how to annotate
โ Repo with folder structure for storage
โ Also on-prem for (commercial) firms
More: https://bit.ly/3L78YI2
๐TORAS: web-based AI-powered, cooperative, annotation platform.
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ SOTA AI tools -> significant speedup
โ "Recipes" to define how to annotate
โ Repo with folder structure for storage
โ Also on-prem for (commercial) firms
More: https://bit.ly/3L78YI2
๐ฅ9๐คฏ2๐1
This media is not supported in your browser
VIEW IN TELEGRAM
๐ฎMAXIM: Multi-Axis MLP for Vision๐ฎ
๐#Google opens MAXIM, a multi-axis MLP for low-level vision
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Denoising, deblurring, dehazing, etc
โ Multi-axis gated MLP, linear complexity
โ Cross gating block, separate features
โ SOTA results on several datasets!
More: https://bit.ly/3Dmp8LI
๐#Google opens MAXIM, a multi-axis MLP for low-level vision
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Denoising, deblurring, dehazing, etc
โ Multi-axis gated MLP, linear complexity
โ Cross gating block, separate features
โ SOTA results on several datasets!
More: https://bit.ly/3Dmp8LI
๐ฅ12โค1๐1
This media is not supported in your browser
VIEW IN TELEGRAM
๐ฅ A Survey on Diffusion Models ๐ฅ
๐A comprehensive review of denoising diffusion models in #computervision ๐คฏ
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Overview on diffusion models
โ Hot trend for the generative AI
โ A multi-perspective categorization
โ Current limitations / new directions
More: https://bit.ly/3RYG5zP
๐A comprehensive review of denoising diffusion models in #computervision ๐คฏ
๐๐ข๐ ๐ก๐ฅ๐ข๐ ๐ก๐ญ๐ฌ:
โ Overview on diffusion models
โ Hot trend for the generative AI
โ A multi-perspective categorization
โ Current limitations / new directions
More: https://bit.ly/3RYG5zP
โค5๐3๐ฅ1