Title: Results remain unchanged after code modifications despite prints confirming execution

[LIU-coder79]

Description
I encountered a behavior where evaluation metrics (IDF1) stopped updating after certain code changes, even when reverting all modifications. The issue persists until the project is fully redeployed.

Steps to Reproduce

1. Deployed the original code → obtained baseline IDF1: 76.692.
2. Modified only sam2tracker.py and run.py in several experimental steps — results updated normally initially.
3. After one specific change, the IDF1 became stuck at 75.318 in all subsequent runs.
4. Even after removing all modifications (or commenting them out), the result remained 75.318.
5. Cleared output caches — no effect.
6. Added print() statements to confirm code execution — prints showed modifications were executed, but the final metric remained unchanged.
7. Only after completely redeploying the project did the results return to normal.

Expected Behavior
Evaluation results should reflect code changes, especially when modifications are reverted or adjusted.

Current Behavior
IDF1 is fixed at 75.318 regardless of code changes, until a full redeploy is performed.

Information Requested

• Are there other caches or state files (e.g., in /tmp, ~/.cache, or inside virtual environment) that should be cleared between runs?
• Any known behavior where certain modules or variables retain state across script executions?

Additional Context
The issue appeared only after several valid iterative changes, suggesting something in the runtime environment or internal state became “stuck.” Full redeployment resolves it, but this is not practical during development.

[JanFMeier]

Hey, thank you for your interest in this project. Can you verify that the evaluation function is reading from the correct output folder? If tracking and evaluation are decoupled, it's possible the evaluation is picking up stale results from a previous run.

Are there other caches or state files (e.g., in /tmp, ~/.cache, or inside virtual environment) that should be cleared between runs?

The only cache in the pipeline is for the zero-shot detector detections. This is intentional so you only need to run the detector once and can then quickly iterate on tracker modifications. If one of your changes affected the detections and you later reverted without rerunning the detector, the cached (modified) detections would still be used, which could explain the persistent difference in IDF1.

Any known behavior where certain modules or variables retain state across script executions?

I haven't encountered this during development and testing personally. Given that a full redeploy fixes it, my best guess is that it's one of the above two issues rather than something in the Python runtime itself.

Could you check whether deleting the cached detections and rerunning from scratch (without a full redeploy) resolves it? That would help narrow down the cause.