Shape Registration in the Time of Transformers
Authors: Giovanni Trappolini, Luca Cosmo, Luca Moschella, Riccardo Marin, Simone Melzi, Emanuele Rodolà
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Qualitative and quantitative evaluations demonstrate that our pipeline outperforms state-of-the-art methods for deformable and unordered 3D data registration on different datasets and scenarios. We evaluate the effectiveness of our architecture on a number of challenges. We begin by analyzing the key components of our model, motivating our architectural choices through ablation studies. Finally, we present our results in the context of matching, registration, and inter class registration. |
| Researcher Affiliation | Academia | Giovanni Trappolini Department of Computer Engineering Sapienza University of Rome giovanni.trappolini@uniroma1.it Luca Cosmo DAIS Ca Foscari University of Venice luca.cosmo@unive.it Luca Moschella Department of Computer Science Sapienza University of Rome luca.moschella@uniroma1.it Riccardo Marin Department of Computer Science Sapienza University of Rome marin@di.uniroma1.it Simone Melzi Department of Computer Science Sapienza University of Rome simone.melzi@uniroma1.it Emanuele Rodolà Department of Computer Science Sapienza University of Rome emanuele.rodola@uniroma1.it |
| Pseudocode | No | The paper describes the architecture and method in text and with diagrams (Figure 2), but does not include pseudocode or explicitly labeled algorithm blocks. |
| Open Source Code | Yes | All code and data is publicly available 1. 1https://github.com/Giovanni TRA/transmatching |
| Open Datasets | Yes | For all our experiments in the humans domain, we trained our method on the same shapes from the SURREAL dataset [55] used in [34]. It consists of 10000 point clouds for training. We also trained our model on the Shape Net dataset [8]. A popular dataset to analyze real identities and poses is FAUST [7], which is composed by ten subjects in ten different poses. |
| Dataset Splits | No | The paper describes training and testing sets, but does not explicitly mention a separate validation set or its specific use for hyperparameter tuning or early stopping. For instance, it states 'using the first 1000 shapes for training, and a different set of 2700 shapes for testing' for ablation studies, which is a train/test split, not a train/validation/test split. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., GPU models, CPU types). |
| Software Dependencies | No | The paper mentions using 'Adam optimizer [28]' but does not provide specific version numbers for software libraries or dependencies (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | In our comparisons we train our model for 10000 epochs using Adam optimizer [28]. We use 32 latent probes of dimension 64, and 8 layers for both the encoder and the decoder. By exploiting the potential of this architecture, we can train our model requiring only a sparse set of ground truth correspondences (10 20% of the total points). |