Hacker News Reader

Comment by dmm | Hacker News Reader

I see object detection systems often compared by accuracy percentages but rarely by speed.

I ask because I'm building a surveillance system that uses object detection to answer simple questions, like: "Is there a squirrel on my bird feeder?", "Is my car in the garage?", "Is there a package on my front step?".

Right now I'm playing with Darknet and yolo from pjreddie because he emphasized performance and I'm getting about 69fps on my 1080ti.

Is speed not considered important by researchers at this time?

Is there any system that provides decent performance on CPUs? Dedicating $700 gpus to detect squirrels on the bird feeder is fine for crazy people like me but others may find that excessive.

View on HN

Inference speed and net size is definitely considered important and some researchers are definitely tackling it (e.g https://arxiv.org/abs/1602.07360). However it's more of an engineering area (meaning lots of engineers are working on it) and less of a research area (meaning fewer papers get produced per work) because there's not much in the way of performance-per-gflop benchmarks or whatever.

CPU performance is good-ish when you're doing sequential stuff e.g it's not bad at old-school RNNs, but for image recognition using a cheap graphics unit is still going to be better, and if the question is "can I use cheaper hardware and get good performance" the answer is yes, you just have to do some tricks (e.g distillation).

Plenty of image recognition projects on the cheap raspberry pi and its graphics unit.

Inference Speed is very important. The TensorFlow object detection model zoo lists timings for their models; we picked our model to meet timing demands on a cell phone.

I don't think CPUs are going to give you decent performance unless you microoptimize the inference for the host architecture and even then the gains are marginal.

If you want to avoid expensive GPUs, take a look at things like Intel's Neural Compute Stick, Google's Vision Kit and the Edge TPU. But you can also run models on modestly priced cell phones- we use $100 cell phones to do object detection and tracking, but it's not pretty.

BTW, try SSD, it was even a bit faster than YOLO2.

Is there a Docker makefile available for playing with SSD if you're not working in the domain?

I have some 100Ks of pix I'd like to put through classifiers but could never get YOLO2 properly working in a container, in part because even reading the code I couldn't figure out how to simply point a trained model at a directory of images and get classification out... :/

Definitely PEBKAC but if there are any dead-simple tutorials available I'd love a pointer–classification speed not a factor, it would all be CPU non-real-time and could be leisurely. Just interested in seeing what the state of the art is in classifying a personal image set without training against it..

Hmm https://github.com/knjcode/mxnet-finetuner looks like maybe what I want...

Any comments from people who know what they're doing welcome :)

There are some wrappers around YOLO3 in Python which may help: https://github.com/qqwweee/keras-yolo3

It's not Dockerised though, but you could grab a container with the usual machine learning packages and install that on top.

Also look at NVIDIA Digits (https://github.com/NVIDIA/DIGITS) which has Docker images.

Thanks Josh!

Checkout MobileNetV2 as a backbone, the paper uses it for SSD and I believe weights are release.

Agreed. I wish that more effort would be made towards reducing inference time, since that is the limiting factor in many real-world applications.

Speed is considered a fundamental problem, e.g. the aforementioned YOLO and _Faster_ RCNN.

If you look for "tensorflow raspberry" on a search engine, you will find several pointers about running TF on cheaper hardware. You would probably get about 1 frame per second with it.

You may also look at recent SnapDragon-powered dev boards, which have TF core accelerators, and provide a serious boost of FPS for less than $700.

Also, using a cloud ML service is an option if you have only a few inferences/day to process

Reply on news.ycombinator.com