Talking Papers Podcast
Deep dives into cutting-edge research papers with their authors
About the Talking Papers Podcast
The Talking Papers Podcast is where research meets conversation: deep dives into research papers in computer vision, 3D, machine learning, and AI, with the authors who wrote them. By researchers, for researchers.
Each episode is structured like the paper itself: a TL;DR / abstract to set the stage, then related work, approach, results, conclusions, and future work. We close with a bonus segment called "What did Reviewer 2 say?", where the author shares the candid peer-review story behind the publication.
Guests are PhD students, postdocs, and faculty from leading labs across academia and industry. Episodes are aimed at fellow researchers and graduate students who want the candid version of the work, not a polished press release. I started the show because I wanted the conversations I wished I'd had earlier in my own PhD, and to put a human voice behind the papers that often feel anonymous on arXiv.
Podcast Episodes
Choosing a PhD Advisor: Questions to Ask + Red Flags
2025-02-17
A frank conversation with Derek Liu (returning guest) on what to actually ask a prospective PhD advisor, the red flags worth taking seriously, and the lessons most students learn the hard way.
Listen →
3D Paintbrush: Local Stylization of 3D Shapes with Cascaded Score Distillation
2024-07-12
Talking Papers with Dale Decatur on 3D Paintbrush — local texture editing of 3D meshes from text via Cascaded Score Distillation. Outputs drop straight into standard graphics pipelines. CVPR 2024.
Listen →
3DInAction Explained: Recognizing Human Actions Directly from 3D Point Cloud Sequences (CVPR 2024)
2024-06-03
Talking Papers with myself (host as guest!) on 3DInAction — recognizing human actions directly from 3D point cloud sequences via 't-patches' that move coherently in time. CVPR 2024.
Listen →
Cameras as Rays: Pose Estimation via Ray Diffusion
2024-03-14
Talking Papers with Jason Zhang (CMU) on Cameras as Rays — reformulating sparse-view camera pose estimation as a per-ray prediction problem with a diffusion-style generative head. ICLR 2024.
Listen →
Instant3D: Fast Text-to-3D with Sparse-View Generation and Large Reconstruction Model
2024-02-16
Talking Papers with Jiahao Li (TTIC) on Instant3D — feed-forward text-to-3D that generates assets in ~20 seconds, two orders of magnitude faster than score-distillation. ICLR 2024.
Listen →
Variational Barycentric Coordinates
2023-12-14
Talking Papers with Ana Dodik on Variational Barycentric Coordinates — replacing mesh- and formula-based generalized barycentric coordinates with a neural field, unlocking new objective-function freedom. SIGGRAPH Asia 2023.
Listen →
Reverse Engineering Self-Supervised Learning
2023-11-22
Talking Papers with Ravid Shwartz-Ziv on Reverse Engineering SSL — an empirical look at what self-supervised learning actually learns. Surprise: semantic clustering comes from the regularization term. NeurIPS 2023.
Listen →
Constructive Solid Geometry on Neural Signed Distance Fields
2023-11-09
Talking Papers with Zoë Marschner on Constructive Solid Geometry for Neural SDFs — fixing the 'Pseudo-SDF' problem when applying booleans and CSG operations to learned signed distance fields. SIGGRAPH Asia 2023.
Listen →
HMD-NeMo: Online 3D Avatar Motion Generation From Sparse Observations
2023-11-01
Talking Papers with Sadegh Aliakbarian on HMD-NeMo — online full-body avatar motion generation from sparse head-and-hands HMD signals. Handles hand occlusion that prior work assumed away. ICCV 2023.
Listen →
CC3D: Layout-Conditioned Generation of Compositional 3D Scenes
2023-09-28
Talking Papers with Jeong Joon Park on CC3D — a 3D GAN that synthesizes compositional scenes with multiple objects, conditioned on a 2D semantic layout. Beyond the single-object 3D-GAN limitation. ICCV 2023.
Listen →
NeRF-Det: Learning Geometry-Aware Volumetric Representation for Multi-View 3D Object Detection
2023-09-07
Talking Papers with Chenfeng Xu on NeRF-Det — multi-view 3D object detection that reasons in a NeRF-style volumetric backbone instead of projecting back to 2D. ICCV 2023.
Listen →
🎠 MagicPony: Learning Articulated 3D Animals in the Wild
2023-08-10
Talking Papers with Tomas Jakab on MagicPony — reconstructing articulated 3D animals (shape, texture, lighting, pose) from a single in-the-wild RGB image, no 3D supervision. CVPR 2023.
Listen →
Word-As-Image for Semantic Typography
2023-07-20
Talking Papers with Shir Iluz on Word-As-Image — automatic semantic typography that morphs letter outlines (as Bézier curves) to convey a word's meaning while staying readable. SIGGRAPH 2023 Honorable Mention.
Listen →
Panoptic Lifting for 3D Scene Understanding with Neural Fields
2023-07-10
Talking Papers with Yawar Siddiqui on Panoptic Lifting — 3D-consistent panoptic segmentation from in-the-wild 2D images by lifting masks into a unified neural-field volume. CVPR 2023 highlight.
Listen →
MobileBrick: Building LEGO for 3D Reconstruction on Mobile Devices
2023-06-14
Talking Papers with Kejie Li on MobileBrick — LEGO objects with known per-brick geometry as physical ground truth for evaluating mobile-device 3D reconstruction. 153 captured models. CVPR 2023.
Listen →
Aligning Step-by-Step Instructional Diagrams to Video Demonstrations
2023-05-17
Talking Papers with Jiahao Zhang on the IKEA Assembly in the Wild dataset — aligning step-by-step assembly diagrams to YouTube videos, a less-explored multimodal alignment problem. CVPR 2023.
Listen →
CLIPasso: Semantically-Aware Object Sketching
2023-03-29
Talking Papers with Yael Vinker on CLIPasso — converting photos into minimal semantic sketches by optimizing Bézier curves against CLIP embeddings. SIGGRAPH 2022 Best Paper.
Listen →
INR2Vec: Deep Learning on Implicit Neural Representations of Shapes
2023-03-29
Talking Papers with Luca De Luigi on INR2Vec — treating an Implicit Neural Representation itself as an input signal that downstream networks can consume, no discrete conversion needed. ICLR 2023.
Listen →
Random Walks for Adversarial Meshes
2022-12-14
Talking Papers with Amir Belder on Random Walks for Adversarial Meshes — generating adversarial examples against triangle-mesh classifiers via random-walk surrogates. First general mesh-domain attack. SIGGRAPH 2022.
Listen →
Stochastic Poisson Surface Reconstruction
2022-12-13
Talking Papers with Silvia Sellán on Stochastic Poisson Surface Reconstruction — a Gaussian-Process-based statistical reformulation of PSR that supports posterior uncertainty queries. SIGGRAPH Asia 2022.
Listen →
Beyond Periodicity: Towards a Unifying Framework for Activations in Coordinate-MLPs
2022-12-06
Talking Papers with Sameera Ramasinghe on Beyond Periodicity — a unifying framework for understanding which activation functions actually make coordinate-MLPs work, beyond the SIREN sinusoid story. ECCV 2022 oral.
Listen →
KeypointNeRF: Generalizing Image-based Volumetric Avatars using Relative Spatial Encoding of Keypoints
2022-10-19
Talking Papers with Marko Mihajlovic on KeypointNeRF — generalizable volumetric human avatars from 2–3 RGB images, conditioned on relative spatial encodings of detected keypoints. ECCV 2022.
Listen →
BACON: Band-Limited Coordinate Networks for Multiscale Scene Representation
2022-08-09
Talking Papers with David Lindell on BACON — band-limited coordinate networks with analytically-known Fourier spectra, finally making multiscale neural-field behavior predictable. CVPR 2022.
Listen →
Learning Smooth Neural Functions via Lipschitz Regularization
2022-07-29
Talking Papers with Hsueh-Ti Derek Liu on Lipschitz regularization for neural fields — penalizing an upper bound on the Lipschitz constant to get smooth latent spaces for shape editing. SIGGRAPH 2022.
Listen →
DiGS Explained: Divergence-Guided Implicit Surface Reconstruction from Unoriented Point Clouds (CVPR 2022)
2022-07-18
Talking Papers with Chamin Hewa Koneputugodage on DiGS — divergence-guided implicit surface reconstruction from unoriented point clouds. No normal-vector supervision required. CVPR 2022.
Listen →
ICON: Implicit Clothed humans Obtained from Normals
2022-07-18
Talking Papers with Yuliang Xiu on ICON — animatable clothed-human avatars from in-the-wild images, robust across unconstrained poses via SMPL-X-conditioned local features. CVPR 2022.
Listen →
Neural RGB-D Surface Reconstruction
2022-07-18
Talking Papers with Dejan Azinović on Neural RGB-D Surface Reconstruction — combining NeRF-style implicit fields with depth supervision in a truncated SDF for more accurate, complete reconstructions. CVPR 2022.
Listen →
Panoptic 3D Scene Reconstruction From a Single RGB Image
2022-07-18
Talking Papers with Manuel Dahnert on Panoptic 3D Scene Reconstruction — unifying geometric reconstruction with semantic and instance segmentation from a single RGB image. NeurIPS 2021.
Listen →
SampleNet: Differentiable Point Cloud Sampling
2022-07-18
Talking Papers with Itai Lang on SampleNet — a differentiable, task-aware point-cloud sampling layer that retains downstream performance at 3% of the original points. CVPR 2020.
Listen →
Shape As Points: A Differentiable Poisson Solver
2022-07-18
Talking Papers with Songyou Peng on Shape As Points — a differentiable Poisson Surface Reconstruction layer that bridges explicit point/mesh representations with implicit fields. NeurIPS 2021.
Listen →
VLN BERT: A Recurrent Vision-and-Language BERT for Navigation
2022-02-24
Talking Papers with Yicong Hong on VLN BERT — a time-aware recurrent BERT that finally makes vision-and-language navigation work over partially observable trajectories. CVPR 2021 SOTA.
Listen →
Neural Parts: Learning Expressive 3D Shape Abstractions with Invertible Neural Networks
2022-02-11
Talking Papers with Despoina Paschalidou on Neural Parts — 3D shape abstractions that stay both geometrically accurate and semantically consistent, thanks to Invertible Neural Networks per part. CVPR 2021.
Listen →
Deep Declarative Networks
2022-02-05
Talking Papers with Dylan Campbell on Deep Declarative Networks — neural-network layers defined implicitly as the solution to an optimization problem, with gradients via the implicit function theorem. TPAMI.
Listen →
DORi: Discovering Object Relationships for Moment Localization of a Natural Language Query in a Video
2022-02-05
Talking Papers with Cristian Rodriguez-Opazo on DORi — temporal moment localization in long videos from natural-language queries, using a language-conditioned graph over discovered object relationships. WACV 2021.
Listen →
Talking Papers Podcast – The Beginning
2022-01-15
Announcing the Talking Papers Podcast — a podcast by researchers, for researchers, helping you finally get through that 'to-read' paper pile. The mission, the format, and what's coming.
Listen →Where to Listen and Subscribe
- Podcast apps (Apple Podcasts, Spotify, Overcast, and more): talking.papers.podcast.itzikbs.com
- YouTube channel: subscribe for the video episodes
- Mailing list: get notified when new episodes drop
- Twitter / X: @talking_papers
Frequently Asked Questions
What is the Talking Papers Podcast?
It is an interview podcast where the authors of cutting-edge research papers in computer vision, 3D, machine learning, and AI discuss their own work. Each episode follows the structure of a research paper and closes with a "What did Reviewer 2 say?" segment about the peer-review story behind it. Hosted by Itzik Ben-Shabat.
How is each episode structured?
Each episode follows the structure of a research paper: a TL;DR / abstract to set the stage, then related work, approach, results, conclusions, and future work. Every episode also closes with a bonus "What did Reviewer 2 say?" segment, where the author shares the candid peer-review story behind the publication.
Who hosts the Talking Papers Podcast?
I host it. My name is Itzik Ben-Shabat. I am a researcher working in computer vision and 3D learning; you can read more on the About page or my homepage.
What topics does the podcast cover?
Computer vision, 3D vision and reconstruction, neural rendering and radiance fields, point clouds, generative models, self-supervised learning, machine learning, and adjacent areas of AI. Most guests work at the intersection of geometry, deep learning, and perception.
Who are the guests?
PhD students, postdocs, and faculty from research labs in academia and industry. Guests are typically the lead authors of the paper being discussed, so the conversation is grounded in their own experience writing and publishing the work.
Where can I listen to the Talking Papers Podcast?
On any podcast app via the show page, which lists Apple Podcasts, Spotify, Overcast, and other platforms. The video versions of episodes are on the YouTube channel.