Your target is a forklift near-miss video classifier — which supervised pretraining corpus do you pick?

## Reframe the question The naive framing is 'which source dataset is closest to my data'. The useful framing is: *what did the source labels force the network to learn, and does my target need the same thing?* A supervised pretraining corpus is a specification for a representation, written in the language of its label set. Two corpora of similar size can produce very different backbones because their labels demand different discriminations. ## Axis 1: does the label set demand the right kind of structure? A forklift near-miss is not an appearance category; it is an event defined by trajectories, closing speed, and the spatial relation between a vehicle and a person over a window of time. Ask what a source task's labels make impossible to fake: - **A large human-action video corpus** with a few hundred action classes cannot be solved from a single frame for most of its classes — many actions differ only in motion and in how a person interacts with an object. Training on it forces temporal features and interaction features. That is the structural match. - **A still-image object corpus** forces excellent per-frame appearance features and nothing temporal. Reusing it is not wrong, but you are then responsible for building and training the temporal stage yourself, on your own small labelled set, which is where the difficulty actually lives. The same reasoning generalises across modalities. If the target were a respiratory wheeze and cough classifier over audio, the right family is a large supervised audio-event tagging corpus: separating hundreds of sound-event labels forces spectro-temporal features — onsets, harmonic structure, noise texture — that a clinical model needs and that a corpus of, say, speaker-identity labels would not build in the same way. The pattern is constant: match the *discriminations*, not the class names. ## Axis 2: scale and label granularity Between two structurally suitable sources, prefer the one with more data and finer, more diverse labels. Fine-grained label sets force finer intermediate features because coarse features cannot separate near-identical classes. A large corpus with a shallow, coarse label set can produce a weaker representation than a smaller corpus with a demanding one. Both axes matter, and interviewers like candidates who refuse to reduce it to raw example counts. ## Axis 3: input distribution distance Warehouse footage from fixed cameras is not web video. Expect ceiling or high-corner viewpoints, wide angles and lens distortion, poor and flickering light, motion blur, heavy occlusion, a static background, and mostly small subjects far from the camera. Web action video is hand-held, well-lit, subject-centred and edited. The larger this gap, the shallower the useful reuse and the more of the stack you should expect to retrain. Practical mitigations: augment target training toward the source's statistics or the source's toward yours; insert an intermediate training stage on a larger in-domain corpus of your own footage if you can label even coarse events cheaply; and match the preprocessing conventions of the source model exactly. ## Axis 4: the non-technical constraints a lead owns - **Licensing and provenance.** Some corpora and some published backbones carry terms that do not permit commercial redeployment. This is a real blocker and it is a lead's job to check it before a team builds on the weights. - **Privacy and consent.** A near-miss classifier watches identifiable workers. Source-model choice interacts with what you are allowed to store, and with a works-council or regulatory conversation that is not the model's problem but is the project's. - **Longevity.** Standardising a team on one backbone family means one preprocessing convention, one feature store and one set of downstream heads. That has real compounding value and is worth some accuracy. - **Fairness of the source.** A source corpus skewed in who and what it depicts pushes that skew into the features, and a safety-relevant classifier is exactly the place where that surfaces as unequal error rates. ## Axis 5: decide empirically, cheaply None of the above beats a measurement. Shortlist two or three candidate sources — for example an action-video backbone and a still-image backbone with a temporal stage on top — and run the *same* protocol on each: same target train and validation split, same augmentation, a small sweep over how much of the stack you retrain, per-configuration learning-rate tuning. Judge on a held-out target split with the metric that matches the operating point you care about, which for a safety event is recall at an acceptable false-alarm rate rather than raw accuracy. Two or three GPU-days of comparison settles what a meeting cannot. ## The one-paragraph answer Pick the source whose labels cannot be predicted without the structure your target depends on; for near-miss detection that is motion and human-object interaction, so an action-video corpus rather than a still-image one. Break ties on scale and label granularity, discount for the distance between web footage and fixed overhead warehouse cameras, check licence and privacy constraints before committing the team, and then verify the shortlist empirically on a held-out target split.