[HikeAndMap]

voc_eval() builds a ground-truth annotation cache (annots.pkl) the first time eval runs for an experiment, then trusts it unconditionally on every subsequent eval - there's no check that it still matches the current imagesetfile/dataset. Iterating on a custom dataset (adding images, re-labeling, re-tiling) between training runs without manually deleting <data_dir>/annotations_cache/ causes the next eval to crash with KeyError as soon as it hits an image that wasn't present when the cache was written.

This is very likely the real, undiagnosed cause behind several long-open upstream issues that all report the same symptom without ever tracing it to a stale pickle cache (#1397, #874, #1490).

Fixed by checking every imagesetfile entry is a key in the loaded recs dict; if any are missing, rebuild the cache from the XML annotations instead of raising. Verified against a real dataset that hit this exact crash - eval now rebuilds the cache and completes normally, and stays fast (no rebuild) on subsequent evals once back in sync.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ luxedo
❌ HikeAndMap

---

HikeAndMap seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}