A Dataset-Centric Survey of Video Understanding:
Benchmark Characteristics, Architectural Evolution, and Future Directions

1Griffith University 2University of New South Wales
Dataset-induced representational biases across pretraining datasets

The same MotionFormer architecture develops different attention patterns when pretrained on datasets with different structures: context-heavy scenes, motion-centric transformations, and hand-object interactions.

The showcased videos are from OWM.

Abstract

Research in video understanding has advanced rapidly, driven by increasingly diverse datasets and more powerful model architectures. While existing surveys often organize progress by tasks, benchmarks, or model families, they provide limited insight into why particular architectures emerged and succeeded.

This survey adopts a dataset-centric perspective: dataset structure shapes model design. Motion complexity, temporal span, compositional hierarchy, multi-agent interaction, and multimodal richness impose distinct learning challenges. These pressures naturally give rise to inductive biases for viewpoint robustness, temporal ordering, long-range dependency modeling, relational reasoning, and cross-modal alignment.

Dataset Properties

Motion, duration, interaction, composition, and modality define the learning signal.

Inductive Biases

Models favor different invariances and evidence patterns depending on the data regime.

Architectural Response

Model families can be understood as responses to evolving dataset requirements.

Act 1 — The framework

Five-Dimension Taxonomy Explorer

Every benchmark is decomposed into five structural dimensions, each coded from its own sub-variables under a fixed rubric. Pick an anchor benchmark to see where it lands.

Each axis sums its sub-variables, normalised to 0–4. The codes are reproduced verbatim from the survey’s coding; only the weighting into a scalar is ours.

What the structure looks like in the raw data

UCF101 skateboarding frames above four Diving48 dive categories that differ only in rotation and posture
Motion

Magnitude vs. fine-grained dependence

UCF101 is recognisable from global spatiotemporal patterns. Diving48 categories differ only in body rotation, posture and execution order.

EPIC-KITCHENS-100 egocentric frames showing continuous hand-object interaction
Relational

Egocentric hand-object interaction

EPIC-KITCHENS-100: a moving first-person camera and continuous manipulation make human-object relations task-critical.

Charades-Ego paired third-person and egocentric clips of everyday activities
Viewpoint

Paired ego / third-person

Charades-Ego records the same overlapping everyday actions from both viewpoints, making viewpoint variability an explicit variable.

Kinetics-400 label taxonomy tree grouping dribbling, dunking and shooting under ball sports
Compositional

Semantic hierarchy

Kinetics-400 organises labels taxonomically. The hierarchy structures the label space, not the timeline.

Breakfast dataset sequence: take bowl, pour cereals, pour milk, stir cereals
Compositional

Procedural hierarchy

Breakfast decomposes one activity into ordered sub-actions. This one introduces temporal dependencies between constituents.

MPII Cooking 2 third-person static camera frames of a cooking procedure
Temporal

Long procedural span

MPII Cooking 2 holds the camera still and lets the activity run long, isolating temporal span from viewpoint change.

The hook

WebVid-10M and Ego4D are both “video + text”. WebVid is coded C/w/s: coarse alignment, weak pairing, sparse grounding. Ego4D is coded D/v/d: dense alignment, verified pairing, dense grounding. Ten million weakly captioned clips score 1.3 / 4 on the multimodal axis; Ego4D scores 4.0. Modality availability is not multimodal structure — which is exactly why a single “multimodal” label cannot predict what a model has to learn.
Act 2 — The qualitative evidence

Structure Leaves a Trace in Attention

If dataset structure really is the driver, then holding the architecture fixed and changing only the pretraining corpus should change what the model looks at. It does. The attention maps below are from identical architectures that differ only in the dataset they were pretrained on.

MotionFormer

Dataset-induced representational biases (final block)

Write video original sampled frames

Original Frames

MotionFormer Kinetics-400 write video block 12 attention GIF

Kinetics-400

MotionFormer Kinetics-600 write video block 12 attention GIF

Kinetics-600

MotionFormer Something-Something V2 write video block 12 attention GIF

Something-Something V2

MotionFormer EPIC-KITCHENS write video block 12 attention GIF

EPIC-KITCHENS-100

Dataset-induced biases across network depth (pre-training on Something-Something V2)

TimeSformer

Dataset-induced representational biases (final block)

Pour video original sampled frames

Original Frames

TimeSformer Kinetics-400 pour video block 12 attention GIF

Kinetics-400

TimeSformer Kinetics-600 pour video block 12 attention GIF

Kinetics-600

TimeSformer Something-Something V2 pour video block 12 attention GIF

Something-Something V2

Dataset-induced biases across network depth (pre-training on Something-Something V2)

Playground

Choose two model/pretrain/video groups. Every block is shown in order.

Act 3 — The historical and quantitative evidence

Benchmark ↔ Architecture Co-evolution

Datasets above the axis, architecture families below. Hover or tap a benchmark to trace what it is associated with — and how that association is supported.

I — family introduced on this benchmark E — commonly evaluated on it R — retrospective, not evidence of causation

Not just a narrative: matched comparisons

Reported accuracies aggregate publications that differ in pretraining corpus, capacity and evaluation protocol, so they can only be read descriptively. These two cases hold one group of factors roughly fixed while another varies. They are associational, not controlled; the residual confounds are stated.