Opens profile photo
Follow
Yun-Chun Chen
@ycchen918
CS PhD student . Affiliated with and . Research scientist intern . Prev research intern .
Toronto, Ontarioyunchunchen.github.ioJoined April 2019

Yun-Chun Chen’s Tweets

I just discovered 'torchview', a cool new python library to visualize the graph of a neural network in Torch. It seems a lot better than previous existing libraries, it shows you all the important details without the fluff.
5
474
We will be presenting our paper today from 11 am to 1 pm CST at Hall J #1024. Come talk to us to learn about the Breaking Bad dataset. breaking-bad-dataset.github.io
Image
Quote Tweet
Image
I’m excited to share our @NeurIPSConf paper: Breaking Bad: A Dataset for Geometric Fracture and Reassembly. This is joint work with @sellan_s, @Dazitu_616, @animesh_garg, and @_AlecJacobson. Paper: arxiv.org/abs/2210.11463 Project page: breaking-bad-dataset.github.io (1/12)
Show this thread
3
📢📢📢 Excited to announce “𝐧𝐞𝐫𝐟𝟐𝐧𝐞𝐫𝐟: 𝐏𝐚𝐢𝐫𝐰𝐢𝐬𝐞 𝐑𝐞𝐠𝐢𝐬𝐭𝐫𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐍𝐞𝐮𝐫𝐚𝐥 𝐑𝐚𝐝𝐢𝐚𝐧𝐜𝐞 𝐅𝐢𝐞𝐥𝐝𝐬” → PDF: arxiv.org/abs/2211.01600 → Homepage: nerf2nerf.github.io → Code: github.com/nerf2nerf/nerf → Video: youtu.be/S071rGezdNM
2
233
Show this thread
A very interesting paper! Wonder how this paper compares with the Neural Motion Fields paper (arxiv.org/abs/2206.14854) by , , , Dieter Fox, and myself.
Quote Tweet
Introducing Neural Grasp Distance Fields, which represent 6-DOF grasps as @neural_fields, enabling joint grasp and motion planning. sites.google.com/view/neural-gr Prior works use discrete grasp sets & multi-stage pipelines. w/@davheld, @_kainoa_, @mhmukadam (@MetaAI, @CMU_Robotics)
Show this thread
Embedded video
GIF
1
5
Show this thread
Geometric Reasoning is at the core of multiple problems in robotics and vision - assembly, archaeology, and medicine It takes Neural Shape Mating to extremes with shattered objects Check out our #Neurips2022 paper "Breaking Bad" 🧵👇
Quote Tweet
I’m excited to share our @NeurIPSConf paper: Breaking Bad: A Dataset for Geometric Fracture and Reassembly. This is joint work with @sellan_s, @Dazitu_616, @animesh_garg, and @_AlecJacobson. Paper: arxiv.org/abs/2210.11463 Project page: breaking-bad-dataset.github.io (1/12)
Show this thread
Image
43
An unforgettable collaboration with wonderful people. This is also my first (co-)first author paper published at top conferences. Come and check our dataset!
Quote Tweet
I’m excited to share our @NeurIPSConf paper: Breaking Bad: A Dataset for Geometric Fracture and Reassembly. This is joint work with @sellan_s, @Dazitu_616, @animesh_garg, and @_AlecJacobson. Paper: arxiv.org/abs/2210.11463 Project page: breaking-bad-dataset.github.io (1/12)
Show this thread
Image
43
This was a really fun collaboration to work on, and a very poetic one: Using our capacity for breaking things to learn how to put them back together. It was also a great pleasure to work with incredibly hard working students like and . Off to New Orleans!
Quote Tweet
I’m excited to share our @NeurIPSConf paper: Breaking Bad: A Dataset for Geometric Fracture and Reassembly. This is joint work with @sellan_s, @Dazitu_616, @animesh_garg, and @_AlecJacobson. Paper: arxiv.org/abs/2210.11463 Project page: breaking-bad-dataset.github.io (1/12)
Show this thread
Image
15
* Why do part assembly methods fail? * Semantic assembly methods leverage part priors by learning global features. However, each fracture does not have a well-defined semantic meaning. Assembling fractures has to rely on learning local features for surface matching. (10/12)
Image
1
1
Show this thread
* How well do baselines perform? * We evaluate 3 baselines on the everyday subset. DGL, a state-of-the-art part assembly method, performs the best both quantitatively and qualitatively. All methods fail drastically in estimating poses and are far from solving the task. (9/12)
Image
1
2
Show this thread
* Baselines * We benchmark 3 state-of-the-art shape assembly deep learning algorithms on Breaking Bad under various settings. Here shows an overview of a typical vision-based shape assembly pipeline. The model takes as input point clouds and predicts respective poses. (8/12)
Image
1
2
Show this thread
* Dataset gallery * Breaking Bad contains 1 million fractured objects, which can be used for various applications, including artifacts for archaeology, everyday objects for vision and robotics, and other objects for video gaming and example-based fracture simulation. (7/12)
Image
1
2
Show this thread
* How do we generate fractured objects? * We adopt a physically based algorithm from Sellán et al. that efficiently generates fractures. For details of fracture simulation, see twitter.com/sellan_s/statu (5/12)
Embedded video
GIF
Quote Tweet
Image
I am beyond excited to finally share that our newest paper, "Breaking Good: Fracture Modes for Realtime Destruction" will be presented at SIGGRAPH Asia 2022. Video: youtu.be/0k_tEk34nJQ Paper: silviasellan.com/pdf/papers/fra Code: github.com/sgsellan/fract 🧵👇
Show this thread
1
2
Show this thread
* Semantic ≠ geometric * Objects in existing datasets are decomposed in a semantically consistent way. However, objects that break naturally due to external forces generally do not break into fragments that are semantically well-defined. (4/12)
Image
1
2
Show this thread
* Why a new dataset? * Machine learning approaches for assembly require large-scale datasets of fractured objects. Existing assembly datasets are constructed based on human or automated semantic segmentation. (3/12)
Image
1
4
Show this thread
After failing to keep (low) stipends up to date with inflation, would like to add stress by reduced health care funding. Dear , you want us to speak well of you at the next grad visit day, right? All these great applicants would probably love to know all of this...
Quote Tweet
@uoft is proposing to reduce healthcare coverage for its Teaching Assistants and Course Instructors (Unit 1 @cupe3902). As Acting VP1 I’m able to get the full breakdown of unit 1 benefits usage and I found some interesting trends. 🧵 (1/15)
Show this thread
2
14