SCOPE-4D
A 4D geometric foundation model designed for endoscopic scenes that simultaneously predicts camera parameters, dense geometric structure, and 3D tissue motion trajectories from monocular RGB video via a single forward pass. It integrates a Common-Residual Motion (CRM) mechanism to disentangle camera motion from tissue deformation.
Timeline (last 90 days)
SCOPE-4D is a 4D geometric foundation model specifically designed for endoscopic scenes, capable of simultaneously predicting camera parameters, dense geometric structure, and 3D tissue motion trajectories from monocular RGB video via a single forward inference pass.
The research team newly collected phantom data and a real colonoscopy evaluation set as a validation benchmark for the model in real clinical environments.
SCOPE-4D first performs Geometric Supervised Fine-tuning on SCOPE-5K to learn geometric priors specific to endoscopic scenes, thereby improving camera pose and depth estimation accuracy.
SCOPE-4D introduces a Common-Residual Motion (CRM) mechanism that decomposes motion into a common motion component and a residual motion component to alleviate the difficulty of distinguishing camera motion from tissue motion.
Experiments show that combining CRM with trajectory supervision further improves camera and depth estimation on top of geometric fine-tuning, and endows the model with dense 3D tissue tracking capability.
SCOPE-4D performs robustly across multiple public and newly collected benchmarks in both in-domain and cross-domain scenarios, and shows a clear advantage over existing methods on 3D tracking tasks.
SCOPE-4D demonstrates better stability in long-sequence colon reconstruction.
The research team conducted a blinded user study, the results of which support the model's perceptual quality evaluation of reconstructions on real clinical videos.
Integrating the joint prediction of camera parameters, dense geometry, and 3D trajectories into a single forward inference pass lays the foundation for real-time applications.
All Facts (9)
SCOPE-4D is a 4D geometric foundation model specifically designed for endoscopic scenes, capable of simultaneously predicting camera parameters, dense geometric structure, and 3D tissue motion trajectories from monocular RGB video via a single forward inference pass.
50%UnverifiedThe research team newly collected phantom data and a real colonoscopy evaluation set as a validation benchmark for the model in real clinical environments.
50%UnverifiedSCOPE-4D first performs Geometric Supervised Fine-tuning on SCOPE-5K to learn geometric priors specific to endoscopic scenes, thereby improving camera pose and depth estimation accuracy.
50%UnverifiedSCOPE-4D introduces a Common-Residual Motion (CRM) mechanism that decomposes motion into a common motion component and a residual motion component to alleviate the difficulty of distinguishing camera motion from tissue motion.
50%UnverifiedExperiments show that combining CRM with trajectory supervision further improves camera and depth estimation on top of geometric fine-tuning, and endows the model with dense 3D tissue tracking capability.
50%UnverifiedSCOPE-4D performs robustly across multiple public and newly collected benchmarks in both in-domain and cross-domain scenarios, and shows a clear advantage over existing methods on 3D tracking tasks.
50%UnverifiedSCOPE-4D demonstrates better stability in long-sequence colon reconstruction.
50%UnverifiedThe research team conducted a blinded user study, the results of which support the model's perceptual quality evaluation of reconstructions on real clinical videos.
50%UnverifiedIntegrating the joint prediction of camera parameters, dense geometry, and 3D trajectories into a single forward inference pass lays the foundation for real-time applications.
50%