Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Abstract Rendering: Certified Rendering Under 3D Semantic Uncertainty
Authors: Chenxi Ji, Yangge Li, Xiangru Zhong, Huan Zhang, Sayan Mitra
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate ABSTRACTRENDER on scenes of varying scales and complexities, represented using GAUSSIANSPLAT and NERF. The scenes include: Lego, Chair, and Drums [2], which are single-object scenes on empty backgrounds; Pine Tree [6], a synthetic boulevard scene with trees; Airport [8], a large-scale photorealistic airport environment; Garden [1], a real-world scene; and Airplane, Truck, and Car, containing objects corresponding to CIFAR-10 classes. ... Through experiments on classification (Res Net), object detection (YOLO), and pose estimation (GATENet) tasks, we demonstrate that abstract rendering enables formal certification of downstream models under realistic 3D variations an essential step toward safety-critical vision systems. |
| Researcher Affiliation | Academia | Chenxi Ji , Yangge Li , Xiangru Zhong , Huan Zhang, Sayan Mitra University of Illinois Urbana-Champaign EMAIL |
| Pseudocode | Yes | Algorithm 1 GAUSSIANSPLAT(Sc G, C, u) Algorithm 2 NERF(Sc N, C, u) Algorithm 3 MATRIXINV(X, Xref, k) Algorithm 4 VR-IND(a, c, d) Algorithm 5 SUMCUMPROD(a, c) |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: Yes, we uploaded our code to github and is reproductive. |
| Open Datasets | Yes | We evaluate ABSTRACTRENDER on scenes of varying scales and complexities, represented using GAUSSIANSPLAT and NERF. The scenes include: Lego, Chair, and Drums [2], which are single-object scenes on empty backgrounds; Pine Tree [6], a synthetic boulevard scene with trees; Airport [8], a large-scale photorealistic airport environment; Garden [1], a real-world scene; and Airplane, Truck, and Car, containing objects corresponding to CIFAR-10 classes. |
| Dataset Splits | Yes | The 360 range is partitioned into angular intervals, and ABSTRACTRENDER computes abstract images for each, which are propagated through the classifier via CROWN [10] for set-based certification. ... We assess a target pose estimator built upon Gate Net [47] (fine-tuned on our airplane and truck datasets) by partitioning the camera’s translational perturbation range into intervals. |
| Hardware Specification | Yes | Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: We run our experiment on 32G A100 GPU. Running times are also provide. |
| Software Dependencies | No | Our implementation, ABSTRACTRENDER, targets two state-of-the-art photorealistic scene representations 3D Gaussian Splats and Neural Radiance Fields (Ne RF) and scales to complex scenes with up to 1M Gaussians. ... Our implementation of abstract rendering employs CROWN [10] for computing such bounds. ... GAUSSIANSPLAT reconstructions are generated using Splatfacto [45], while those of NERF are trained with the vanilla NERF algorithm [2], both implemented with standard Nerfstudio settings. |
| Experiment Setup | Yes | We test whether a pretrained CIFAR-10 Res Net [46] maintains correct predictions as the camera orbits the target object azimuthally over 360 at fixed distance and elevation. The 360 range is partitioned into angular intervals... We assess a target pose estimator...by partitioning the camera’s translational perturbation range into intervals. ... to verify whether positional errors between pose estimation and ground truth holds below a given error tolerance (20% divation from ground truth). ... Table 3: ... d: Distance to Object Centroid (m); h: Camera Height to Object’s Horizon Plane (m); Npart: Number of Partitions... Table 4: ... d: Distance to Object Centroid; CPR: Camera Perturbation Range (m); ... Thr: Threshold |