GS-3DORC: Advancing 3D Object Retrieval & Completion with Gaussian Splatting
(Proposal for SHREC 2025)
Embedded Files
Organizers:
- Thien-Phuc Tran*, University of Science, VNU-HCM, Vietnam, ttphuc21@apcs.fitus.edu.vn
- Minh-Quang Nguyen, University of Science, VNU-HCM, Vietnam, nmquang21@apcs.fitus.edu.vn
- Minh-Triet Tran*, University of Science, VNU-HCM, Vietnam, tmtriet@hcmus.edu.vn
- Tam V. Nguyen, University of Dayton, U.S.A, tamnguyen@udayton.edu
- Minh Do, University of Illinois at Urbana-Champaign, U.S.A., minhdo@illinois.edu
β¦
For questions, please contact the organizers at ttphuc21@apcs.fitus.edu.vn.
1. Motivation
3D Gaussian Splatting is a novel and promising approach for 3D scene and object representation, offering efficient rendering and reconstruction capabilities. However, many classical challenges in 3D object and scene retrieval remain unresolved. This challenge seeks to leverage Gaussian Splatting to improve 3D retrieval, reconstruction, and completion tasks, testing model robustness, generalization, and accuracy in various scenarios.
2. Challenge Statement
This AI challenge aims to advance 3D object understanding using Gaussian Splatting. Participants will tackle retrieval and reconstruction tasks under increasing levels of difficulty, pushing their models to generalize across different distributions.
Task 1: 3D Part Retrieval
Objective: Given images of a complete 3D object, retrieve its corresponding parts (3-10 parts) from a database of Gaussian Splatting models.
Challenges:
Shape recognition and segmentation
Accurate retrieval of relevant parts
Controlled conditions for evaluation
Task 2: 3D Part Completion
Objective: Given an incomplete 3D object, retrieve the missing part(s) from a database and predict their exact placement in the original model.
Challenges:
Shape alignment and part fitting
Spatial reasoning and object reconstruction
Handling occlusions and incomplete geometry
Task 3: Generalized 3D Retrieval & Completion
Objective: Extend Task 1/2 to a different test distribution. Models trained on the original dataset must generalize to unseen distributions.
Challenges:
Adaptation to unseen categories and styles
Robustness in retrieval and reconstruction
Evaluating real-world generalization
For Task 3, evaluation will be conducted using 3D Gaussian Splatting models reconstructed from the Objaverse-XL dataset to assess the transferability of learned representations.
3. Dataset & Evaluation
Dataset
All 3D models will be represented in Gaussian Splatting format. The primary dataset for this challenge is ShapeSplatsV1, with additional synthetic datasets generated using DreamGaussian to enhance variety and complexity.
The dataset consists of 500 Gaussian Splatting objects of various classes, which will be released in the following subsets:
Starter Set (Training)
Public Test Set
Private Test Set
Evaluation Metrics
To rigorously assess model performance, we propose the following evaluation criteria:
Retrieval Accuracy: Measures the percentage of correctly retrieved parts for Task 1 and Task 2.
Spatial Alignment Error: Quantifies how accurately retrieved parts are positioned in Task 2.
Generalization Performance: Assesses how well models trained on ShapeSplatsV1 adapt to unseen distributions in Task 3.
4. Conclusion
This challenge will advance research in 3D retrieval, object reconstruction, and scene understanding using Gaussian Splatting. By addressing part retrieval, completion, and generalization, we aim to improve real-world applications in 3D design, virtual environments, and automated content generation. Participants will push the boundaries of multi-modal learning, spatial reasoning, and adaptive 3D representation.
Important Dates
π₯ Registration Opened: 03 April 2025
π Public Data Released: 05 April 2025
π Registration Closed: 12 April 2025
π Private Data Released: 12 April 2025
π’ Notification of Results: 20 April 2025
π Technical Report Submission: 25 April 2025
Page updated
Report abuse