[Q] Was the reported FPS achieved using GPU?

[navervision]

[Q] Was the reported FPS achieved using GPU?

[geonm]

Yes. We used V100 GPUs to get all reported FPS numbers with a native ckpt model not tflite model.

If you use tflite converted models (we call it tflite model here), it will use only CPU and it sometimes shows slower inference time than mobile devices even it’s on a high-performance CPU (Xeon something) because TFLITE OP Kernels are optimized for ARM processors.

We highly recommend to use tflite models for mobile devices.

There are also the native ckpt models in this repository.

Btw, is ssd mobilenet v3 also a tflite converted one?

[ValYouW]

Hi,

Yes, I am using your provided tflite model (thx!), I'll surely test it on my Android.

As for the SSD, yes, I converted it to tflite, the project I am using is this, you can find there the tflite model.
If I remember correctly I took the model from here

Thanks again.

[ValYouW]

@geonm
So I tested on Android, using the M-LSD_320_tiny_fp16 model, on my Pixel 4A 5G I get ~250ms per detection.
On the same device using ssd_mobilenet_v1 I get ~50ms per detection.

If you're interested you can run the code from my repo: https://github.com/ValYouW/tflite-crossplatform (clone repo and open the android folder in Android Studio)

Thx.

P.S.
I also tried using the nnapi delegate but there was no real difference...

[geonm]

Thank you for reporting valuable numbers.

Honestly, our mlsd model has to be slower than ssd_mobilenet_v1.

We adopt the slightly modified mobilenet as a backbone but we designed a decoder part like U-Net which requires more computational cost than SSD and also we employed a larger input image (320x320).

Unlike object detection, the evaluation metric of line segment detection (especially wireframe dataset) mostly depends on how well a model could detect adjacent line segments.

If we adopt SSD as a decoder, the performance would drastically drop.

[ValYouW]

Thank you for the explanation, really nice work!