PaperlayerView exact revision on arXiv

Source-linked AI summary

From Wrecks to Wisdom: Recovering Crash Mechanics from Real-World Multi-View Photos

Ondřej Valach, Václav Diviš, Ivan Gruber

arXiv:2609.39486v1cs.CV

TL;DR

Crash-mechanics information is often missing or delayed in standard records, despite the availability of post-crash photos. This paper estimates CDC descriptors and reconstructed ∆V components from multi-view photo sets; its selected joint-training pipeline reduces principal-direction-of-force angular error by 30.1% versus the single-task reference.

  • Problem

    Crash-mechanics descriptors may be missing or delayed in standard records, although they are important inputs to downstream risk modeling and triage and post-crash photos are often available early.

    Source: §1 · p. 2

  • Method

    The paper predicts CDC-derived deformation descriptors and directional ∆V components from multi-view photo sets, using a shared visual architecture and comparing single-task with joint training.

    Source: §1 · p. 3

  • Results

    The selected joint-training pipeline reduced principal direction of force angular error by 30.1%, from 20.1° to 14.05°, while improving several other targets under the compared configurations.

    Source: §6 · p. 11

  • Takeaways & Limitations

    The CISS-based protocol and reference results provide a basis for future image-based crash-mechanics estimation and multimodal fusion with structured crash metadata.

    Source: §1 · p. 3

    Visual observability varies across targets, with ∆V least directly observable; joint-training gains are not uniform across descriptors.

    Source: §4.1 · p. 7, §5.2 · p. 11

Abstract

from arXiv · show

Estimating accident mechanics from real-world crashes is important for vehicle-safety analysis, injury modeling, crash-severity prediction, and operational workflows such as insurance claim triage. In standard crash records, key metadata such as impact configuration, principal direction of force, and change in velocity ($\Delta V$) may be missing, delayed, or corrupted, while post-crash photographs are widely available and contain rich visual evidence of deformation. We study how much crash-mechanics information can be recovered directly from vehicle photos when structured signals are absent. We formulate crash understanding as supervised prediction from per-case multi-view photo sets. Targets include six Collision Deformation Classification (CDC) descriptors and the longitudinal and lateral components of reconstructed $\Delta V$. Each photo is encoded by a shared visual backbone, and the resulting view-level features are fused into a case-level representation from which target-specific heads predict crash descriptors. Using 15.2k training cases from the Crash Investigation Sampling System, drawn from about 1.5M photos before filtering, together with 1.15k validation and 1.15k test cases, we define an evaluation protocol for vision-based crash descriptor estimation from incomplete multi-view evidence. Post-crash imagery alone provides usable signal for several non-trivial crash-mechanics descriptors, while weakly observable and long-tailed targets remain challenging. Within the compared training regimes, the selected joint-training recipe reduces mean absolute angular error for principal direction of force from 20.1 to 14.05 degrees and longitudinal $\Delta V$ MAE from 8.04 to 7.45 km/h. Our work provides a reference point for future multimodal fusion with structured crash metadata.

1 Introduction

Crash-mechanics descriptors matter for safety and downstream modeling, but are often unavailable in standard crash records. This study tests whether multi-view post-crash photos can support their estimation and establishes a reference protocol for doing so.

1 Introduction

  • Crash-mechanics information can be absent or incomplete in standard records, and many lower-severity collisions go unreported.Parameters describing real-world crash mechanics are often not directly measurable; the passage notes that some collisions are unreported because they do not require a tow.Source: §1 · p. 2
  • Crash descriptors are important inputs to downstream risk modeling and triage, but are typically available only after Event Data Recorder extraction or crash reconstruction.Photos are often among the earliest and most consistently available information channels.Source: §1 · p. 2
  • The study predicts six CDC-derived deformation descriptors and directional ∆V components from post-crash photo sets, comparing separate single-task training with joint training.The targets capture complementary aspects of collision geometry and dynamics.Source: §1 · p. 3
  • The paper establishes a CISS-based protocol for estimating crash descriptors from naturally incomplete multi-view post-crash photos, with preprocessing and case construction for canonical view slots and missing labels.It reports single-task and joint-training reference results for CDC-derived descriptors and reconstructed ∆V components.Source: §1 · p. 3

2 Related Work

Prior crash-severity research generally uses structured records and treats crash mechanics as available predictors, while image-based methods have addressed narrower settings. This paper instead examines multi-target recovery from real-world, naturally incomplete photo sets.

2 Related Work

  • Most crash-severity and injury pipelines use structured records, treating ∆V, principal direction of force, and peak acceleration as predictors rather than quantities inferred from images.These mechanics descriptors are treated as informative signals for injury estimation and severity modeling.Source: §2 · p. 3
  • Statistical reconstruction can infer unobserved quantities from structured records, but crash-mechanics variables in such settings are inputs when available, not prediction targets.The paper instead investigates recovery directly from photos when structured signals are absent, delayed, or incomplete.Source: §2 · p. 3
  • Prior image-based studies typically predict narrow outputs, use curated single-image settings, or supplement images with synthetic data or vehicle metadata.The closest cited studies use real and synthetic images or a curated frontal-only subset with vehicle metadata; this paper offers a real-world multi-view reference point without claiming direct comparability.Source: §2 · p. 4

3 Dataset

The dataset pairs real-world post-crash photos with investigator-coded CDC and reconstructed ∆V targets. Cases are organized into disjoint partitions and represented with up to nine canonical viewpoints despite missing views and labels.

3 Dataset

  • The CISS-based evaluation setup contains 17.5k crash cases, drawn from approximately 1.5M photos before filtering.CISS was selected for being openly available, large, standardized, and richly documented with post-crash imagery; GIDAS was excluded due to access constraints and RASSI due to differences in road and traffic environment.Source: §3 · p. 4
  • For each case, targets include an investigator-coded eight-character CDC code and reconstructed WinSMASH ∆V components, from which six CDC descriptors are derived.The descriptors are principal direction of force, deformation plane, longitudinal/lateral and vertical/lateral crush zones, damage distribution, and deformation extent.Source: §3 · p. 4
  • Cases are split at the crash-case level into disjoint train, validation, and test partitions of 15,200, 1,150, and 1,150 cases, respectively.Because target distributions are strongly long-tailed, the study reports accuracy and tail-sensitive metrics such as macro-F1.Source: §3 · p. 4
  • The model input is an ordered set of up to nine canonical viewpoints, while crash photos may include uninformative close-ups, missing views, and incomplete target annotations.Figure 2 illustrates representative cases in the fixed multi-view format.Source: §3 · pp. 4–5, Figure 2 · p. 5

4 Methods

The methods encode ordered crash-photo views with a shared visual backbone and fuse them into a case representation for descriptor prediction. They address incomplete, imbalanced real-world data and compare single-task and joint-training regimes, with task-specific formulations for different targets.

4 Methods

  • A shared SwinV2 backbone encodes post-crash views, followed by case-level multi-view fusion; the backbone is chosen for hierarchical local-to-global features, efficient window attention, and positional interpolation.These properties are intended to suit subtle or localized evidence while keeping computation feasible at the study's image resolution.Source: §4 · p. 5

4.1 Preprocessing and Data Challenges

  • The preprocessing challenges include irrelevant close-ups, missing viewpoints, incomplete annotations, and strongly imbalanced labels; the study addresses them with filtering, fixed view slots, view-drop augmentation, masked losses, and imbalance-aware sampling.The observability of targets differs: impact plane and principal direction of force are often supported by visible damage layout, whereas crush zones, deformation extent, and especially reconstructed ∆V are less directly observable.Source: §4.1 · pp. 6–7

4.2 Method Setup

  • The method setup compares single-task models, which learn one descriptor per instance, with a joint model using multiple task-specific heads over a shared trunk.Preliminary ablations favored independently encoded SwinV2 view tokens fused with a lightweight transformer.Source: §4.2 · p. 7

4.3 Descriptor Targets and Task Formulations

  • Principal direction of force is treated as 12-class circular prediction with Gaussian-smoothed targets and circular-distance angular error; other categorical descriptors use cross-entropy, while ordinal deformation extent uses smoothed 9-way classification.For ∆V, longitudinal and lateral components are jointly regressed with SmoothL1 loss; only the 62% of cases with valid labels are used because ∆V is missing for 38%.Source: §4.3 · pp. 7–8

4.4 Single-task

  • Single-task training learns one descriptor per model instance using a task-specific head and task-specific sampling, skipping cases without that target.All single-task models share the same case representation, backbone, and fusion module.Source: §4.4 · p. 8

4.5 Multi-task

  • The joint-training configuration shares the backbone and fusion module across targets, with a shared pre-head and task-specific heads; task-specific loss masking handles missing labels.Its comparison with single-task references is not a perfectly matched ablation: initialization, loss balancing, sampling, and optimization schedule also differ.Source: §4.5 · p. 8

5 Experiments and Results

The experiments show that post-crash photos support some crash descriptors better than others, while the selected joint-training pipeline improves several targets but not uniformly. Results are reported at validation-selected checkpoints and should be read as comparisons between practical training regimes, not as an isolated test of multi-task supervision.

5 Experiments and Results

  • Accuracy is interpreted alongside macro-F1 where applicable because the most frequent class sets an always-majority accuracy baseline under matched label distributions.Small differences between selected training regimes are treated cautiously.Source: §5 · p. 9, Table 1 · p. 9
  • Test results use validation-selected checkpoints, and the single-task and joint models use different selected configurations, so their comparison does not isolate multi-task supervision.The joint pipeline also differs in initialization, sampling, and loss balancing.Source: §5 · p. 9

5.1 Single-task

  • Single-task models learned impact plane and DoF well and showed usable signal for longitudinal/lateral zone, but long-tailed labels challenged vertical/lateral zone and damage distribution.Deformation extent had substantially higher near-miss agreement under Acc@±1 than exact prediction; longitudinal and lateral ΔV MAE were 8.04 km/h and 5.31 km/h, respectively.Source: §5.1 · pp. 9–10

5.2 Multi-task

  • 20.1°→14.05° DoF angular error was the largest gain in the selected joint-training comparison, which also improved deformation extent and both ΔV components.The comparison reflects selected training configurations rather than an isolated ablation of multi-task supervision.Source: §5.2 · p. 10, Table 2 · p. 10
  • Joint training was not uniformly better: impact-plane accuracy changed marginally, while long./lat. zone macro-F1 and damage-distribution accuracy declined.The authors suggest shared sampling and loss balancing may under-serve targets with different long-tail structure.Source: §5.2 · p. 11

6 Conclusion

The study establishes an evaluation protocol for crash-descriptor estimation from naturally incomplete, real-world multi-view photos and finds usable image signal for several descriptors and reconstructed ΔV components. Some targets remain limited by weak observability and class imbalance, motivating multimodal follow-up.

6 Conclusion

  • The CISS-based protocol evaluates crash-descriptor estimation from naturally incomplete, real-world multi-view post-crash photo sets.The selected joint-training pipeline improved several targets, most notably reducing DoF angular error by 30.1% from 20.1° to 14.05°.Source: §6 · p. 11
  • Weak observability and strong class imbalance leave some descriptors difficult to recover, so image-based crash understanding remains incomplete in this real-world setting.Source: §6 · p. 11
  • A proposed next step is combining photo evidence with structured crash metadata, with curated better-balanced subsets also suggested for rare descriptors.The suggested metadata include vehicle, occupant, roadway, and environmental factors.Source: §6 · p. 11

1 Preliminary Experiments on Direction of Force (CDC Clock)

Preliminary DoF experiments favored transformer-based multi-view processing over the tested CNN-style approaches, and the fixed ordered-view transformer setup produced the most stable learning and lowest validation error. These explorations informed the final model design.

1 Preliminary Experiments on Direction of Force (CDC Clock)

  • CNN-style EfficientNet-B4 and InternImage-T runs plateaued early, although InternImage-T showed a modest improvement when views were channel-concatenated.Source: §1 · p. 13

1.1 Per-Setup Description (P1-P6)

  • P1 used EfficientNet-B4 with replicated random views, concatenation, and circular regression, while P2 used InternImage-T with the same view construction and circular smoothed classification.Both configurations showed limited learning progress or unstable training.Source: §1.1 · p. 14
  • P3 fed channel-stacked images to InternImage-T in one pass, yielding modest improvement over P2 but an early validation plateau.Source: §1.1 · p. 14
  • P4 mean-pooled randomly permuted SwinV2 views and learned more consistently than CNN baselines, while P5 concatenated replicated random-K SwinV2 views and achieved lower angular error than P4.The P4–P5 comparison also changes view construction.Source: §1.1 · pp. 14–15

1 Preliminary Experiments on Direction of Force (CDC Clock)

  • SwinV2-base with fixed ordered views and transformer fusion had the most stable learning and lowest validation error in the preliminary experiments.The setup reached 21.8° AngErr; this preliminary ablation value is not the best single-task DoF result.Source: §1.1 · p. 15

1.1 Per-Setup Description (P1-P6)

  • P6 used a fixed ordered nine-slot input with padding, transformer fusion, and circular smoothed classification; its validation result motivated the final design.The ablations cover a compact subset of the exploratory model-development experiments.Source: §1.1 · p. 15

2 Dataset Curation, Preprocessing, and Evaluation Details

The section describes how CISS cases and incomplete photo sets are curated and organized for evaluation, with filtering and label handling addressing practical data challenges. Fixed-slot view organization improves DoF prediction, while missing or weak evidence and long-tailed labels remain constraints.

2.1 Crash Case Acquisition and Curation Pipeline

  • CISS cases are cleaned and filtered through metadata-based image selection, vehicle detection, and wheel-dominance removal before forming curated crash-image sets.The pipeline removes unusable cases and visually uninformative images while retaining crash-relevant exterior vehicle views.Source: §2.1 · pp. 15–16, Figure 1 · p. 16

2.2 Preprocessing and Data Challenges

  • The input data include uninformative close-ups, missing viewpoints, incomplete target labels, and strongly imbalanced classes, requiring preprocessing and training accommodations.Wheel close-ups can carry little case-level deformation information; unavailable viewpoints remain empty, and losses are masked for targets lacking labels.Source: §2.2 · p. 16

2.3 Wheel-Detector Filtering Details

  • A RetinaNet wheel detector trained on CAWDEC removes wheel-dominated images before slot assignment while retaining broader side views relevant to case-level deformation.The detector addresses close-ups that CISS metadata alone cannot identify as uninformative for global deformation reasoning.Source: §2.3 · p. 17, Figure 2 · p. 17

2.4 9-Slot Construction

  • Cases use nine canonical viewpoint slots, with one seeded representative image per slot; metadata-based organization means this is not fully unconstrained image-set learning.Unavailable views are padded, and deterministic seeded assignment is used for validation and test cases.Source: §2.4 · pp. 17–18
  • Both fixed-slot variants substantially reduce DoF angular error versus unordered random views, and auto-orientation matches metadata slots at reported precision.The comparison averages three runs; metadata slots are used in the main experiments to isolate descriptor learning from viewpoint-estimation noise.Source: Table 2 · p. 18, §2.4 · p. 18

2.5 View-Drop Augmentation and Missing Labels

  • Training randomly drops 0–6 available views, while missing-label samples are skipped per single task or contribute only to valid heads under masked multi-task losses.At inference, the model receives only available images, without synthetic view dropping.Source: §2.5 · p. 18

2.6 Target Coverage and Label Imbalance

  • CDC-derived categories are strongly long-tailed, with frontal and central configurations more frequent than rarer patterns, and weak or missing damage evidence limits fine-grained estimates.Failure cases include occluded primary damage, subtle deformation despite reconstructed ∆V, rare CDC classes, and close-up or partial photo sets.Source: §2.6 · pp. 18–19, Figure 3 · p. 19

3 Training and Optimization Details

The section presents per-target single-task training and a shared multi-task recipe, including their distinct sampling strategies. Because the selected configurations also differ in initialization, loss balancing, sampling, and optimization schedule, their comparison is not an isolated test of joint supervision.

3.1 Single-Task Training Details

  • Single-task models share the case representation, backbone, and fusion module, while using target-specific heads and sampling choices to address label imbalance.Inverse-frequency sampling and stratified accumulation windows ensure a minimum number of rare samples per effective update in the described scheme.Source: §3.1 · p. 20

3.2 Multi-Task Training Details

  • Joint training uses a shared pre-head and task-specific heads, with task-aware sampling that weights cases by their rarest available categorical target.GradNorm and Dynamic Weight Averaging are enabled from epoch 6; the recipe does not guarantee balanced batches for every target.Source: §3.2 · pp. 20–21
  • The selected single-task and joint-training comparison is not a matched ablation of multi-task supervision because initialization, loss balancing, sampling, and optimization schedule also differ.Both regimes share the case representation, backbone, and fusion module, but single-task hyperparameters are tuned per target.Source: §3.2 · p. 20, Table 3 · p. 21

4 Reference Points to Prior Image-Based Work

The section positions the study relative to prior image-based crash work while emphasizing differences in targets, data curation, crash-mode coverage, and input modalities. Those differences make the reported prior-work figures contextual reference points rather than direct baselines.

4 Reference Points to Prior Image-Based Work

  • Prior image-based crash studies use different targets, curation, crash-mode coverage, or inputs, so their reported numbers are contextual references rather than direct baselines.Silver et al. combine real and synthetic images, while Hasija et al. use real-world cases with vehicle metadata.Source: §4 · pp. 21–22
  • Impact-plane comparisons contrast prior single-image front/back prediction with this benchmark’s real-world multi-view cases, natural viewpoint missingness, and four-plane coverage.Because label spaces, curation protocols, and inputs differ, the numbers are not evidence of performance parity.Source: Table 4 · p. 22, §4 · p. 22
  • Prior ∆V results are included to contextualize the real-world image-based setting, not as direct baselines, given differences in data and evaluation.The study frames its results as a practical reference point and multimodal fusion with structured metadata as a next step.Source: §4 · p. 22

Related papers

Loading 2609.39486v1…