face detection performance

Gilles Caulier caulier.gilles at gmail.com
Wed Mar 29 09:36:57 BST 2023

Hi all,

Some information about performance of face detections and OpenCV
library: In all bundles (MAcOS, Windows, and AppImage), the GPU usage
in OpenCV is disabled for stability reasons. We received a huge list
of bug reports about this topic in the past, so i disabled all
optimizations in OpenCV.

But here, I compiled OpenCV with optimizations. digiKam does not crash
using GPU and Multicore (i9-32 cores + 64G RAM + Nvidia T400 + NVME).

So I think a try with OpenCV GPU optimizations can be published in the future.

Other information : the Yolo v3  DNN model is not optimized for GPU
usage. As I can see with a student who will work on the Face
management this summer (and also in a new feature to detect objects,
monuments, places, plants, etc.), the usage of the new Yolo v4 will
introduce the GPU optimizations in the neural network.

We haven't yet experimented with something in this way. But as I can
see in NVIDIA cards specifications, a T400/T600/T1000 and certainly
more expensive video cards, includes hardware DNN cores where models
can be loaded to do the job directly on GPU, not the CPU.

Also in my office we have some project to analyze experimental
fast-real-time measurements data using GPU and neural network. So the
story will continue...


Gilles Caulier

Le mer. 29 mars 2023 à 10:06, Thomas <sdktda at gmail.com> a écrit :
> On 2023-03-29 09:54, frederic chaume wrote:
> > Is it the performance I should expect ? Is there a way to improve the
> > performance ?
> >
> >
> I will be following this thread as I also notice the face
> detection/recognition being surprisingly slow. I have an order of
> magnitude more pictures in my collection than you. And my collection is
> located on a network share. So we are talking days or weeks to run a
> full scan.
> I also noticed that most cores go underutilized. Same with network
> throughput - it is quite low. I have thought that it might be the NAS
> server being the bottleneck but I have done careful analysis of that and
> the disks are not busy in it and not obvious bottleneck was found on the
> machine hosting the network shares.
> I have several clients using my collection. Each have their own digikam
> deployment and local digikam db files stored on fast local SSDs.
> All clients are fast computers with modern hardware. Example spec is
> core i7 with 64 GB RAM and SSD. Other clients are 2022 MacBooks with M1
> CPUs.
> --
> Mvh
> Thomas

More information about the Digikam-users mailing list