Camera

Hardware

We use the USB-3 Basler a2A1920-160umBAS with Basler Lens C11-2520-12M-P f25mm to generate the vide recordings. A profile file for this camera can be used to replicate the camera configuration. This camera can deliver up to 160 FPS at 2.3MP resolution. We record at 150 FPS and crop a square ROI out of the sensor feed, which is rectangular (1920x1200). The square is rougly 1000x1000 (1MP)

Software

We use scicam to interact with the cameras and ffmpeg to save the video data in a memory efficient format.

scicam is a wrapper library around pypylon, the Python module produced by Basler (the camera supplier) to drive Basler cameras.

To install it: pip install scicam==0.1.3

scicam is called as follows

CUDA_VISIBLE_DEVICES=1; scicam --setup FlyHostel3 --chunk-duration 300 --start-time   1682604000 --sensor 192.168.0.103:9000  --X 1

CUDA_VISIBLE_DEVICES=1 will make the GPU with ID 1 (as reported by nvidia-smi) to be in charge of efficiently encoding the video to .mp4.

nvidia-smi

Wed May  3 08:32:33 2023       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 525.105.17   Driver Version: 525.105.17   CUDA Version: 12.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA GeForce ...  Off  | 00000000:15:00.0 Off |                  N/A |
| 30%   27C    P8    15W / 350W |      6MiB / 24576MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
|   1  NVIDIA RTX A4000    Off  | 00000000:21:00.0  On |                  Off |
| 41%   30C    P8     9W / 140W |    488MiB / 16376MiB |      3%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

This output shows CUDA_VISIBLE_DEVICES=1 will put the NVIDIA RTX A4000 in charge. The advantage of this GPU is that its only limit on how many videos it can encode is its memory, and not the number of slots. The NVIDIA GeForce card actually has more RAM, but it can only take 3 videos simultaneously, even if they only took say half the memory. This is summarised in this table https://developer.nvidia.com/video-encode-and-decode-gpu-support-matrix-new#Encoder

  • --sensor contains the URL (IP address and port) of a sensor server that returns a JSON with fields temperature humidity, light and camera_light, as explained in the sensor page.
  • --chunk-duration 300 is a number in seconds that states how long each imgstore chunk will be (300 seconds = 5 mins).
  • --select-rois prompts a GUI that allows users to select the ROI for each camera feed, that way non relevant parts of the feed are not saved, which saves space.
  • --setup FlyHostel3 specifies which of the 4 setups to run. Options are FlyHostel1 FlyHostel2 FlyHostel3 FlyHostel4.
  • --X 1 number of animals in the experiment. This feature allows scicam to create the experiment data in the right X folder (a folder where all experiments have the same number of animals)
  • --start-time 1682604000 timestamp in seconds when the experiment started, in this case 2023-04-27 14:00:00.

Scicam uses a derived version of imgstore available under https://github.com/shaliulab/imgstore. See more in imgstore