Two original clips were shared from RCSI. Click the link below to see the clip.
Dark cycle Sz: Sent to microsevenX on 06-21-2024 Dark cycle, file size: 1.06 GB, video hours: 7.01.56 hours, video features: 1280x720p and ten fps—scanning time: 14 hours. The scanning per frame time is 49 ms per frame. The total is about 400000 more frames.
Light cycle Sz: Sent to microsevenX on 06-21-2024 Light cycle Sz, file size: 764.3 MB, video hours: 2.16.18 hours, video features: 1280x720p and ten fps. Scanning time: 7 hours. The scanning per frame time is 49 ms per frame. The total is about 200000 more frames.
The original video is 720p and ten frames per second.
The report of Dark cycle Sz from microsevenX:
The console screenshot:
The report of Light cycle Sz from microsevenX:
The console screenshot:
The detect seizure model used the miceNJ dataset from Rutgers cameras and its environment. The annotation is miceNJ. If a seizure is detected, the future microsevenX system can find multiple mice on the same screen with various annotations.
Here are some questions to discuss:
1 The camera and the cage environment:
a. Lights setup/background
b. The camera lens/view of the angle
c. 1080p and the frame rate (if needed, WDR [wide dynamic range] and night vision sensor)
2. Database:
In the current development setup, the database is using Microsoft Excel. We are collecting the frame coordinate parameters. The dark cycle sz is more than 400000 frames, and the light sz is more than 200000 frames. Excel provides a maximum of 65536 rows.
We will install and set up the MS SQL database in the microsevenX system. It can help the report save data for repeating the result, and the report can be archived. We only scan the clip once. MS SQL is one of the options. We will consider the SQL database to be cross-platform in different OSs. There are many cloud databases offered from Google Vertex, Oracle, and Amazon AWS. It will cost and have data hosted somewhere.
We can not save the coordinate parameter frame data using Excel. And the report can’t be saved. See the following error console screen:
3. Detect seizure model training computer needs a Nvidia CUDA-compatible computer. If we have 1,000 training frames, it will take 9 hours on a regular computer, but it takes about 10 minutes on a Nvidia CUDA-compatible computer: more trained dataset and more worth and accuracy.
4. The coordinate curve shows some of the frames detected via the miceNJ dataset:
5. Dataset:
The dataset should be built for RCSI mice and cages, which do not exist in the miceNJ (Rutgers) dataset. The miceNJ can not detect the mice in the RCSI cages, and the Redbox is not shown. See the following screenshots.
6. False detected:
These are falsely detected via the miceNJ dataset. We will modify the algorithm to correct the false detection.
