Deleting predictions removes labels from unrelated labeled video #2472
Description
Bug description
When correcting predictions on new data and adding corrected frames into the training dataset, deleting predictions unexpectedly removed labels from another fully annotated video.
After updated to v1.5.2, I performed inference on a full video, I manually corrected mispredicted frames and used “Add current frame” to include them into the training dataset. Then, I went to the Labels tab and used “Delete All Predictions”.
This action not only deleted the predictions from that video, but also deleted my existing labels from another manually annotated video, which should never be affected by “Delete All Predictions”. This caused loss of labels.
Additionally, inference speed in the current version is extremely slow on my system (only ~2–4 FPS), compared to ~15–30 FPS in v1.4.1a2 using the same hardware and workflow.
Expected behaviour
- “Delete All Predictions” should only remove prediction items for the currently loaded video, not erase valid labels from other labeled videos.
- Inference speed should be significantly higher than 2–4 FPS on high-end GPU hardware.
Actual behaviour
- “Delete All Predictions” removed predictions for the current video and also deleted labels for another video, causing data loss.
- Inference runs at only ~2–4 FPS.
Your personal set up
OS: Ubuntu 22.04.2 LTS (GNU/Linux 5.15.0-112-generic x86_64) inside an LXD container
Hardware: 2 × AMD EPYC 7642 2.3GHz CPUs, 3 × NVIDIA A6000 GPUs, 512 GB RAM
SLEAP version: v1.5.2
Installation method: Installed with uv tool
How to reproduce
- Run inference on a video with a trained model.
- Inspect predictions and manually correct frames with errors.
- Click “Add current frame” to add corrected frames into the training dataset.
- Navigate to Labels → click “Delete All Predictions”.
- Labeled instances from another previously annotated video disappear.