When a concrete or rock cylinder is crushed in a Split Hopkinson Pressure Bar, it fails within a few hundred microseconds. The Split Hopkinson Pressure Bar (SHPB) is a pair of long steel rods that squeeze a small specimen with a stress wave. Strain gauges on the rods record that wave, and the record gives stress and strain averaged over the whole specimen. Experimenters film the test with high-speed cameras and run digital image correlation (DIC), which tracks a painted speckle pattern to measure surface deformation. In brittle materials this pipeline breaks at exactly the moment that matters. Spalled chips and dust fly toward the lens, hide the still-intact load-bearing core, and cast shadows over it. Correlation fails, and the regions that fail are masked out. What is left is either a strain map with holes over the fracture zone or nothing at all. The physics that would explain why fibres toughen concrete, or why water-saturated rock shatters more finely, is then inferred from sieving the debris after the test.
Computer vision has the same structure under a different name. Virtual try-on and human–object rendering struggle with "self-occlusion artifacts", where one surface hides another and naive methods either bleed the hidden texture through or erase the occluder. We propose treating a fracturing SHPB specimen the same way. A deformable continuum (the core) is partly hidden by a growing population of nearly rigid small objects (the fragments). Both are reconstructed jointly from two or three synchronized high-speed cameras as separate layers of 3D Gaussians, which are small coloured blobs optimized to reproduce the images. This gives two quantities nobody currently measures during the event. The first is the deformation of the core, including behind debris. The second is the time-resolved kinetic energy carried away by fragments. The second can be checked against the dissipated energy computed independently from the bar gauges.
Core Claim
The idea borrows from GASPACHO GASPACHO: Gaussian Splatting for Controllable Humans and Objects. That paper reconstructs interacting humans and objects as decoupled sets of 3D Gaussians. It also uses an occlusion-aware photometric loss, which excludes pixels hidden by one entity from the gradient of the other. It reports PSNR 28.31 on DNA-Rendering, against 26.49 for Animatable Gaussians and 21.13 for HUGS. Its stated limitation is that the object must be a single rigid body. In fracture diagnostics that limitation reverses usefully. Each ejected fragment is close to rigid, while the core is the deforming part. Our representation is therefore a deformable core layer plus a variable number of rigid fragment layers.
A detachment rule moves Gaussians from the core into a new fragment cluster. The rule fires when a connected group of core Gaussians starts moving as a rigid body and separates from its neighbours. The claim is that this layered model, constrained by multiple views and by the bar-gauge record, recovers core displacement under occluders and fragment kinetics more accurately than masked stereo-DIC. It also yields an energy budget that closes against the bar measurements.
The closest existing experimental work is A Test Method for Finding Early Dynamic Fracture of Rock: Using DIC and YOLOv5. It pairs DIC with an object detector to flag where cracks appear. Detection answers where a crack is. We aim to measure what moves underneath the debris and how much energy leaves with it.
Research Program
The physical motivation comes from three SHPB studies that end at post-mortem evidence. Wen et al. show in Enhanced Dynamic Compressive Behavior of Rubberized Concrete with Steel–Glass Fibers that hybrid steel–glass fibres raise dynamic strength from 40.73–61.29 MPa to 60.25–101.86 MPa. In the same work the fineness modulus of the debris rises from 3.20–3.33 to 3.73–4.20, but fragmentation is assessed only by sieving afterwards. Jian et al., in Energy Dissipation and Damage Evolution of Water-Saturated Skarn Under Impact Loading, split the bar-measured energy into input, elastic, and dissipated parts. They find that saturated rock fragments more finely and accumulates damage faster in the plastic stage. They cannot say how much of the dissipated energy is fragment kinetic energy rather than new crack surface. The friction and inertia analysis in Numerical and Theoretical Analysis of the Inertia Effects and Interfacial Friction in SHPB Test Systems shows that friction can inflate transmitted stress by 9–19%. It also explicitly lacks experimental kinematics to validate its simulations. These papers establish the gap. None supplies a method.
The method borrows three pieces from vision. GASPACHO supplies the layered Gaussian representation and the occlusion-aware loss: a core pixel contributes to the core's gradient only where the rendered fragment layer leaves it visible. The second piece follows the argument in BooW-VTON: Boosting In-the-Wild Virtual Try-On via Mask-Free Pseudo Data Training and the erasure failure acknowledged in Stable VITON, namely that hard binary masks destroy information at boundaries. Accordingly, the visibility used in our loss is soft and rendered from the current reconstruction, not thresholded from correlation scores. The third piece is a smoothness prior on core displacement in the spirit of attention-as-flow regularization from Learning Flow Fields in Attention for Controllable Person Image Generation. It lets core motion behind an occluder be interpolated in time and space from visible neighbours, with a per-Gaussian uncertainty that grows with occlusion duration.
We then add one mechanics-specific constraint. The axially averaged core strain must match the strain reduced from the reflected bar wave, $\epsilon(t) = -\frac{2c_b}{L_s}\int_0^t \epsilon_r(\tau)\,d\tau$, where $c_b$ is the bar wave speed, $L_s$ the specimen length, and $\epsilon_r$ the reflected strain pulse. This match is enforced only before the onset of fracture, while the one-dimensional assumption holds.
We do not adopt 3D mesh templates as in 3DV-TON. No parametric shape prior exists for a crumbling cylinder. Latency, around 30 s per sequence there, is irrelevant here, because reconstruction runs after the test. The practical constraint is data. High-speed cameras at roughly 10⁵–10⁶ frames per second record only tens of low-resolution frames per test, so the model must be fitted per test with strong priors rather than trained on large data.
Synthetic ground truth comes from explicit simulation. We use ABAQUS/Explicit fracture runs of the kind in the Pei et al. study, and particle DEM runs of brittle cylinders as in Guoxin Li's DEM work. These are rendered with speckle texture, realistic lighting, motion blur, and camera noise at real frame counts and resolutions. Physical tests use concrete mixes mirroring Wen et al. (S0G0 and S1.2G0.4) and dry versus saturated rock cylinders mirroring Jian et al. Each test is filmed by three synchronized cameras, two used for fitting and one held out.
Baselines
The naive reference point is the bar record alone. Stress, strain, and dissipated energy come from the one-dimensional formulas, and nothing is said about core deformation or fragment energy. Every optical method must at least agree with it before fracture.
The strongest existing competitor is high-speed stereo-DIC with standard subset correlation thresholding, as used in the stereo-DIC literature represented by Automated Camera Exposure Control for Accuracy-Enhanced Stereo-Digital Image Correlation Measurement. Subsets whose correlation drops below a threshold are discarded, and the remaining field is median-filtered. This is what an experimental mechanics lab would actually do. It is run on the same two fitting cameras and frames, with its subset size, step, and threshold tuned on synthetic sequences using the same tuning budget as our method's hyperparameters. No reported result found on Lacuna for stereo-DIC accuracy under fragment occlusion in SHPB tests, so we measure this baseline ourselves.
To beat it, our method must achieve lower displacement error on core regions occluded for at least three frames in synthetic data. It must not be worse on visible regions. It must also score higher on held-out camera prediction in real tests. A second baseline adds fragment handling to DIC: background subtraction segments moving debris, DIC runs on the rest, and fragments are tracked as blobs by a per-frame centroid tracker to estimate kinetic energy.
The key ablation keeps our full layered model but replaces the rendered soft visibility with hard masks derived from correlation failure. It isolates whether the occlusion-aware layering, not merely multi-view Gaussians, drives any gain. A second ablation removes the detachment rule so that all motion is treated as a single deformable layer. A third removes the bar-strain constraint, to show how much the mechanics anchor matters.
Why not just…?
Why not just run stereo-DIC and discard occluded subsets? On visible core regions it may well be as accurate as our method, and we expect that outcome. The difference lies in what it cannot report. The holes it leaves sit over the fracture zone, and it measures nothing about fragments. Our synthetic benchmark separates errors into visible and occluded regions, so a tie on visible regions and a win only on occluded ones is the expected and honest result. If our method does not win on occluded regions, the interpolation prior adds nothing and the proposal reduces to a fragment tracker.
Why not just segment the debris and track it with a simple blob tracker? This is the real threat to the kinetic-energy half of the claim. Fragments are small, move nearly ballistically, and could plausibly be tracked by centroid motion alone. Its weaknesses are clusters of overlapping fragments, fragments still attached at their first frames, and mass estimation from a 2D silhouette. Our 3D layered model attacks all three: multi-view volume gives mass, and detachment timing comes from the rigid-motion test. We will report both estimates against the ground-truth fragment kinetic energy in simulation. If the blob tracker comes within the noise of our method, we will say so and keep only the core-reconstruction contribution.
Why not just correct the bar data analytically or with simulation? Pei et al. show that friction and inertia corrections depend on the assumed constitutive law, impact velocity, and evolving geometry, and their own analysis lacks experimental kinematics for validation. Bar records give boundary averages and cannot separate crack-surface energy from fragment kinetic energy. That separation is the scientific question for the Jian et al. and Wen et al. materials.
Why not just add more cameras or mirrors? More views genuinely reduce occlusion, and our method uses multiple views. But debris clouds in front of a cracking specimen hide it from every nearby angle, and frame budgets limit how many high-speed cameras a lab can run. The held-out camera experiment shows directly how much a third view adds when it is used for fitting versus for testing.
Proposed Experiments
On synthetic fracture sequences, we compare core displacement error on visible versus occluded regions. We also compare fragment kinetic energy error and the timing of fragment detachment across all methods and ablations, sweeping debris density and frame rate.
On real tests, the decisive checks are threefold. The first is held-out camera photometric error on core pixels. The second is agreement between optically averaged core strain and bar-reduced strain before fracture. The third is energy closure: the reconstructed fragment kinetic energy must be positive and smaller than the bar-derived dissipated energy, $U_d = U_i - U_r - U_t$ (the input, reflected, and transmitted wave energies). It must also stay consistent across repeat tests at the same impact pressure.
The scientific payoff is tested by two predictions. If fibres arrest cracks as Wen et al. argue, the hybrid-fibre mix should send a smaller fraction of $U_d$ into fragment kinetic energy than plain rubberized concrete. If water accelerates plastic-stage damage as Jian et al. infer, saturated rock should show earlier fragment detachment and a sharper rise in ejecta energy. A method that cannot resolve either difference beyond test-to-test scatter has failed, whatever its image metrics.
References
- GASPACHO: Gaussian Splatting for Controllable Humans and Objects
- A Test Method for Finding Early Dynamic Fracture of Rock: Using DIC and YOLOv5
- Enhanced Dynamic Compressive Behavior of Rubberized Concrete with Steel–Glass Fibers
- Energy Dissipation and Damage Evolution of Water-Saturated Skarn Under Impact Loading
- Numerical and Theoretical Analysis of the Inertia Effects and Interfacial Friction in SHPB Test Systems
- BooW-VTON: Boosting In-the-Wild Virtual Try-On via Mask-Free Pseudo Data Training
- Stable VITON: Learning Semantic Correspondence with Latent Diffusion Model for Virtual Try-On
- Learning Flow Fields in Attention for Controllable Person Image Generation
- 3DV-TON: Textured 3D-Guided Consistent Video Try-on via Diffusion Models
- Propiedades mecánicas del hormigón corroído por la lluvia ácida con base en el elemento discreto de partículas (Editado por el Dr. Kimon Christianis, University of Patras, Greece)
- Automated Camera Exposure Control for Accuracy-Enhanced Stereo-Digital Image Correlation Measurement