Shashank Agnihotri

Today, from 16:00-18:00 at MICCAI 2026, we will be presenting our work: Faithful, Interpretable Chest X-ray Diagnosis with Artifact-free B-cos Networks. Shreyash Arya, Marcel Kleinmann, Bernt Schiele, Margret Keuper. Poster number: M-PM-142 Please come by to discuss further. :)

I will be presenting our paper, "Failure or Drift? Evaluating Monocular SLAM under Synthetic and Real-World Corruptions", at the 3rd NeuSLAM Workshop at ECCV 2026 in Malmö on 8 September! The paper asks a simple but important question: when visual conditions become difficult, does a SLAM system fail clearly, or does it keep tracking while quietly accumulating drift? We compare ORB-SLAM2, DPVO, and DROID-SLAM across synthetic and real-world adverse conditions, and find that the choice and realism of the corruption can change which system appears most robust. I really enjoyed working on this with Abhay Thomas and @margret_keuper (@dwsunima and @cvml_mpiinf), and I’m looking forward to discussing it with the NeuSLAM community! Paper: arxiv.org/pdf/2608.30690 Workshop: sites.google.com/view/neusla… Happy to chat over coffee about this and other topics at the European Conference on Computer Vision; feel free to hit me up! :) #ECCV2026 #NeuSLAM #SLAM #ComputerVision #Robotics

I’ve been told I should share more of my work, so here goes! At @icmlconf 2026 in Seoul, I presented a spotlight talk on "Images as Tables: In-Context Learning with TabPFN for Low-Data Detection of AI-Generated Images" at the Foundation Models for Structured Data Workshop. The central question is slightly unusual: can images be treated as tables? We turn frozen image representations into structured rows and use TabPFN for in-context real/fake classification. This provides a lightweight way to adapt to new image generators when only a small number of labelled examples are available. I really enjoyed the questions and discussions! And I am grateful to the workshop organizers for the opportunity to present it. 😄 🙌 Paper 📜: openreview.net/pdf?id=6uEcEW…

Strong performance on clean benchmarks is not enough. Vision systems must also survive blur, noise, compression, weather, and other real-world shifts. Introducing RoCo-Spring, an official @NeurIPSConf 2026 Competition Track challenge on robust dense correspondence. Compete in: Optical Flow | Stereo Matching | Scene Flow | Exploration We have made participation easy: register your team, download the data, clone the PTLFlow-based starter devkit, run the pipelines, and submit. 🌐 roco-spring.github.io/ 💻 github.com/hmorimitsu/roco-s… Questions? [email protected] or DM me. #NeurIPS2026 #ComputerVision #Robustness Organizers: @shashankska, @JennySchmalfuss, Victor Oei, Katrin Bauer, Henrique Morimitsu, Andres Bruhn, @margret_keuper Affiliations: @dwsunima, @nvidia, @cvml_mpiinf, @Uni_Stuttgart, and University of Science and Technology Beijing.

🧠 Introducing RoCo-Spring: The Robust Correspondence Challenge at @NeurIPSConf 2026 Dense correspondence methods: optical flow, stereo matching, and scene flow have made remarkable progress on clean benchmarks. However, as FlowBench (Agnihotri et al., 2025) shows (below), robustness and generalization to real-world distribution shifts have remained largely stagnant. Methods that excel on clean data often fail under noise, blur, adverse weather, or compression artifacts - conditions that are unavoidable in autonomous driving, robotics, and augmented reality. RoCo-Spring turns this gap into a community challenge: Can we build methods that are both accurate on clean data AND robust under realistic visual shifts? We measure the best possible trade-off between: 📐 ID performance on Spring 🌧️ OOD performance on RobustSpring …across 3 downstream tasks: 🔹 Optical flow 🔹 Stereo matching 🔹 Scene flow The challenge features both a Competition Track, with leaderboards for each task, and an Exploration Track for rigorous scientific studies into why current methods succeed or fail in robustness evaluations. 📅 Workshop in-person at #NeurIPS2026 Starting kit, codebase, and full details coming soon! Organizers: Jenny Schmalfuss, NVIDIA Victor Oei, University of Stuttgart Katrin Bauer, University of Stuttgart Andres Bruhn, University of Stuttgart Shashank Agnihotri, University of Mannheim Margret Keuper, University of Mannheim and MPI-INF Henrique Morimitsu, University of Science and Technology Beijing #RoCoSpring #OpticalFlow #StereoMatching #SceneFlow #RobustVision #ComputerVision

Two years in a row! 🥳 Hopefully, again next year. 🤞

Presenting the following two works at @iclr_conf 2026: GeoDiv on Thu, Apr 23, 2026 • 10:30 AM – 1:00 PM, Pavilion 4 P4-4117 Project page: abhipsabasu.github.io/geodiv… RobustSpring on Sat, Apr 25, 2026 • 3:15 PM – 5:45 PM, Pavilion 4 P4-3703 Project page: spring-benchmark.org/

Our latest submission to @sigbovik : TBME beats SchmidhuberAI, as theoretically and empirically proven by us! To read what we could release about the paper, visit: data.dws.informatik.uni-mann… No questions/feedback please! #SIGBOVIK #AI #LLM #VLM #(AM)²-M

Built a PyTorch convolution visualizer for teaching bachelor’s/master’s students. Something that instructors can point students to when they need a clearer understanding of how modern deep learning libraries actually compute convolution. github.com/shashankskagnihot…

Currently underway in Peridot 202 #AAAI2026 Singapore Expo at the AIR-FM workshop!

The keynotes at the AIR-FM Workshop are currently underway! Join us in Peridot 202 for the some interesting talks on Reliability of LLMs and VLMs in the Real World! #AAAI2026 @aaai #VLMs #LLMs #Hallucinations #Reliability #RealWorld