The cheapest question comes before accuracy
A team with a folder of domain images and no ground truth has two expensive options in front of it: label the pool, or fine-tune a model. Both are worth deciding on evidence, and there is a much cheaper question that comes first: does a ready-made detector react to this material at all?
That is a screening question, not an evaluation. It needs no labels and no training step — one inference pass over the raw images, and the predicted boxes are the answer. What it buys is direction: whether to spend on annotation, on a different pretrained candidate, or on adaptation.
The pool is real material with its labels held back
The images are the public Mini Traffic Detection pool — real traffic scenes originally assembled and fully annotated for detection benchmarking: vehicles, pedestrians, and roadside signage such as traffic lights and stop signs.
None of that annotation is used here. The pool is handed to the detector as a plain folder of photographs, which is the situation this screen is for. That matters for fairness in both directions: the run gets no help from labels that a real team would not have, and because those labels do exist upstream, the same pool can later back a measured evaluation once screening says it is worth one.
| Setting | Value |
|---|---|
| Model | EdgeCrafter ecdet_s, pretrained weights, no fine-tuning |
| Task | Detection |
| Confidence threshold | 0.4 |
| Batch size | 4 |
| Class filter | None — every class the model knows |
Every image came back annotated
All 226 images were processed and returned with the detector’s own boxes; none were skipped. That sounds like plumbing, and it is exactly what a screen needs to establish first — a model that fails on half a pool’s resolutions, aspect ratios, or color profiles is a finding of its own, before anything about object quality comes up.
The answer arrives in the model’s vocabulary, not yours
The more useful finding is what the categories say. A pretrained detector does not answer in the
taxonomy of the pool; it answers in the public class list it was trained on. So the predictions on
these traffic scenes contain car, person, truck, bicycle, motorcycle, traffic light, and
stop sign — and also dog, handbag, chair, and potted plant.
Both halves of that are information. The first says the model already covers most of what a traffic pool cares about, so a fine-tune would be an adaptation rather than a cold start. The second says the raw output is not a drop-in substitute for domain labels: it has to be mapped, filtered, or narrowed before it means anything in the team’s own schema.
Three scenes worth looking at
Three examples make the review concrete without reproducing the full 226-image gallery. Each box is one prediction, colored by predicted class; the exports carry no text labels and no ground truth overlay, because this run had none.
Together they cover the range that matters to a screening decision: a dominant near object and few-pixel far ones in the same frame, a cluttered scene with many overlapping subjects, and a fine-grained roadside object. Boxes appear across all three — which is the directional signal, not a score.
What this cannot tell you
With no reference set in the run, there is no precision, no recall, no mAP, and no accuracy claim of any kind. The boxes show where the model responded, not how often it was right.
The harder gap is misses: an object the detector ignored leaves nothing behind to look at. The 0.4 confidence threshold also shapes the review directly — a lower one would surface more tentative boxes and a higher one would hide borderline ones, and none of that changes the model, only what a reviewer sees. And three convincing images say nothing about coverage across lighting, weather, camera angles, or rare classes.
The defensible conclusion stays narrow: on this pool, this detector produces behavior worth taking to the next step.
Turning the screen into a decision
The output of a screen is a choice about where the next money goes:
- Label a representative subset when boxes land on the regions reviewers care about. That creates the ground truth a real evaluation needs — and the number that this run cannot produce.
- Try another pretrained candidate when the boxes rarely correspond to the intended objects or scene conditions. Screening is cheap enough to run several times.
- Fine-tune on domain examples when the model clearly understands the broad categories but answers in the wrong vocabulary or misses the domain’s specifics.
Whichever path follows, the import and the inference ran as one reproducible sequence, so the boxes being reviewed are tied to this exact pool and this exact model version rather than to a notebook and a screenshot. Point the same run at a new batch — a different camera, city, or season — or swap the pretrained checkpoint, and the comparison is against a recorded baseline instead of a memory of one.