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..
Asymmetric Dual Self-Distillation for 3D Self-Supervised Representation Learning
Authors: Remco Leijenaar, Hamidreza Kasaei
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Asym DSD achieves state-of-the-art results on Scan Object NN (90.53%) and further improves to 93.72% when pretrained on 930k shapes, surpassing prior methods. We evaluate Asym DSD through extensive experiments across 3D recognition, few-shot classification, and part segmentation, including studies on scalability and ablations. Results to these experiments are shown in Table 1. |
| Researcher Affiliation | Academia | Remco F. Leijenaar Hamidreza Kasaei Department of AI, University of Groningen, The Netherlands |
| Pseudocode | No | The paper describes methods in paragraph text and visually with diagrams (e.g., Figure 2 for overview, Figure 8 for the processing pipeline, Figure 9 for building blocks, Figure 13 for different predictor designs) but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Code available at https://github.com/RFLeijenaar/Asym DSD |
| Open Datasets | Yes | Objaverse 1.0 Scan Object NN Image Net-1k Objaverse-XL MS COCO CIFAR-10 MNIST Shape Net Model Net40. We follow the common SSRL pre-training protocol involving Shape Net Core [36], consisting of 41,952 CAD models across 55 categories. On Scan Object NN [37], a real-world scanned object dataset... On Model Net40 [38], a clean synthetic dataset... We also evaluate semantic segmentation on Shape Net-Part [70]. |
| Dataset Splits | Yes | On Scan Object NN [37], a real-world scanned object dataset, Asym DSD achieves 90.53% (+7.0%) on the hardest PB_T50_RS split, surpassing all prior methods with a standard transformer by +3.6%. We evaluate few-shot performance on Model Net40 following the m-way, n-shot protocol of [69]. In particular, 10 instances are randomly sampled per category, and the remaining instances are added to the test set. |
| Hardware Specification | Yes | With a single RTX 4090, this takes roughly 18 hours to complete. Asym DSD-S* was trained for 100 epoch on a batch size of 128, totaling 727 k optimization steps, in roughly 100 hours on a single A100 GPU. |
| Software Dependencies | No | The paper mentions techniques like 'Precision FP16 mixed [78]', 'Optimizer Adam W [79]', 'Normalization RMSNorm [80]', and 'Activation GELU [81]', which refer to training methods, optimizers, and activation functions. However, it does not explicitly provide specific software dependencies or library versions (e.g., PyTorch version, Python version, CUDA version) used for implementation. |
| Experiment Setup | Yes | We follow the common SSRL pre-training protocol involving Shape Net Core [36], consisting of 41,952 CAD models across 55 categories. Pre-training is run for 300 epochs using Adam W, with a cosine learning rate schedule peaking at 5.0e-4, and a cosine EMA decay increasing from 0.995 to 1.0 during training. For additional details, we refer the reader to Appendix B.1, which includes Table 7 with data pre-processing parameters, training hyperparameters, and model hyperparameters. |