STUDY IN PREPARATION
Truck detection on a new camera
How well does a detector trained on selected port cameras work on a camera it has never seen? How much do a few labeled images from that camera help?
The pilot has not produced cross-camera results yet. The label budgets below are proposed.
Task
Find visible road trucks with bounding boxes and count them in each image. Human reviewers define the reference boxes and record uncertain cases. Still images do not measure arrivals, waiting time, or port throughput.
Data
Audit each camera for a useful truck view, image quality, date coverage, and usage rights. Freeze the chosen image keys, hashes, camera IDs, capture times, and exclusions in a versioned manifest before evaluation.
Test settings
- Train the base detector on two cameras at one port.
- Test it on two other cameras at the same port, using frames from a later week chosen in advance.
- A camera at a different port is outside this pilot. That needs a separate study.
Local labels
Start with no training images from the new camera. Then compare proposed budgets of 10, 50, and 100 human-labeled images from it against the same number of extra labeled images from the training cameras. Keep the target-camera test images fixed and out of training. The pilot will determine feasible final budgets.
Readout
Report box precision and recall, average precision, count error, and labeling time. Show results for each new camera separately, with uncertainty, missed trucks, false alarms, and cases where adaptation does not help.