홍석쓰 블로그

Jitendra lunch group meeting; source: (Israel / The Technion)Probability Theory 101What is the "convolution property of a PDF (probability density function)"?For two independent random variables, the PDF of their sum is the convolution of their individual PDFs. What’s the difference between "uncorrelation" and "independence" in expectation?x1​ and x2​ are uncorrelated if and only if the expectat..

What is KV caching? KV caching is specifically related to the auto-regressive approach of a transformer decoder. In a transformer decoder, it attends to the past and current tokens, but not to future tokens. At each time step, the transformer repeatedly calculates the attention scores between the query and the key, and computes the values by multiplying the scores with the previously computed va..

This always causes problems...1) fail to build mmcv from the source, which was recommended by the egohumans repo2) Why don't I just use pip and mim to install mmcv, mmdet as usual. 3) The usual mmcv installation failed to import nms from mmcv.ops import nms4) Ok, I found out that there is something called mmcv-full and I needed that for nms import (and potentially other functions that egohumans ..

Tested methods: WHAM, GVHMR, TRACETL;DRThey all give poor results. They don't have ground estimation. I made issues in each repo to clarify whether I am doing something wrong or it is the fundamental limitation of their methods. (Oct 12th, 2024)https://github.com/yohanshin/WHAM/issues/118 Fail to put multiple persons in the same world frame · Issue #118 · yohanshin/WHAMHI @yohanshin @dalgu90 , T..

top left: rgb video for reference \ top right: rectified slam left video input; right video input is omitted \ bottom left: normalized depth for visualization with a mask where approximates the overlapping regions of left and right camera FOV \ bottom right: normalized predicted disparity from RAFT-stereo 1. Preprocess SLAM left-right camera imageshttps://github.com/hongsukchoi/generic_tools/blo..

Input: Two hand keypoints from a SPML-H mesh sequenceOutput: Two Shadow robot hand joints' angles + base poseNo dynamics, just geometry. People say it's easy, but it was freaking difficult for me> Why do I choose this input and output? or what are the technical differences?Options: Humanoid, Bi-manual arms with five finger hands, Flying robot hands (what I did) Why didn't I do humanoid retargeti..

TL;DRHuman hand is a tool that has two functions: sensor and motorSummary Hand Function Categories: The framework divides hand function into four categories along a sensorimotor continuum, ranging from tactile sensing to non-prehensile skilled movements. This structure helps analyze the factors influencing human manual performance.Dual Role of the Hand: The hand serves two main functions—sensory..

What is prehension?"A grasp is every static hand posture with which an object can be held securely with one hand, irrespective of the hand orientation". - Feix et al. (2016, p. 67)"Grip is when you move your arm all around and the object still deosnt' fall out" - Kamakura et a. (2022, p. 30)So "hook" is not a form of grip according to them. depends on what hopok is though... What is "Taxonomy of..

1. Install TurboVNC in the remote server, if it is not installed2. SSH to the remote server normally. don't need -X, -Y, -L, -A. Then run,/opt/TurboVNC/bin/vncserver :5017 -rfbport 5017 -SecurityTypes none  -localhost -vgl -fg -log /dev/stderr   (if you are not sudo, just replace /dev/stderr with /tmp/turbovnc.log)Keep this running.2. Install VNC in your local laptop, or TigerVNC if your laptop ..

HuMoR: 3D Human Motion Model for Robust Pose Estimation원래 내 생각보다 더 대단하고 영향력이 더 있고, 앞으로 더 참고할만한 논문인 것 같음.During training, HuMoR learns a conditional Variational AutoEncoder that estimates different transition distributions per previous timestep’s state x (includes joint location, velocity, etc… ) - See “Section 3 Latent Variable Dynamics Model” In addition to the encoder of VAE that encodes the t..