5,000 Shipping Labels: Build and Evaluate a Batch OCR Pipeline

The challenge

Scale up from a few labels to 5,000. Build a job that can resume after a failure, then see how well it handles different layouts, small text, and noisy images.

Start here

Open the SharePoint folder linked above. label_5000.zip contains 5,000 PNG files, label_1.png through label_5000.png. shipping_labels_1757261530888.json contains the answer records. The archive is about 208 MB compressed; inspect its directory and extract a small pilot batch before expanding everything.

The archive's member order starts with high-numbered files. Numeric filename order, ZIP order, and reference-array order are different concepts. Establish matching rules from the label content.

Try it

1. Create a resumable image-to-JSON or image-to-CSV pipeline with one output per unique input.

2. Recover nested addresses, tracking numbers, dates, and weights without losing postal-code formatting.

3. Measure accuracy separately for noisy and non-noisy labels and identify error patterns.

4. Reproduce the manifest counts and weight totals from the reference after the extraction experiment.

What to build

Build a manifest, a processing-status file, a manually reviewed sample, and an evaluation table containing field matches, the number of fields tested, unresolved records, and clean/noisy breakdowns. State whether you processed all 5,000 images or only a sample.