SDK for running DeepLabCut on a live video stream
This package contains a DeepLabCut inference pipeline for real-time applications that has minimal (software) dependencies. Thus, it is as easy to install as possible (in particular, on atypical systems like NVIDIA Jetson boards).
Performance: If you would like to see estimates on how your model should perform given different video sizes, neural network type, and hardware, please see: https://deeplabcut.github.io/DLC-inferencespeed-benchmark/
If you have different hardware, please consider submitting your results too! https://github.com/DeepLabCut/DLC-inferencespeed-benchmark
What this SDK provides: This package provides a DLCLive
class which enables pose estimation online to provide feedback. This object loads and prepares a DeepLabCut network for inference, and will return the predicted pose for single images.
To perform processing on poses (such as predicting the future pose of an animal given it's current pose, or to trigger external hardware like send TTL pulses to a laser for optogenetic stimulation), this object takes in a Processor
object. Processor objects must contain two methods: process and save.
process
method takes in a pose, performs some processing, and returns processed pose.save
method saves any valuable data created by or used by the processorFor more details and examples, see documentation here.
poetry install deeplabcut-live
, thanks to PR #60.Please see our instruction manual to install on a Windows or Linux machine or on a NVIDIA Jetson Development Board. Note, this code works with tensorflow (TF) 1 or TF 2 models, but TF requires that whatever version you exported your model with, you must import with the same version (i.e., export with TF1.13, then use TF1.13 with DlC-Live; export with TF2.3, then use TF2.3 with DLC-live).
pip install deeplabcut-live
Note, you can then test your installation by running:
dlc-live-test
If installed properly, this script will i) create a temporary folder ii) download the full_dog model from the DeepLabCut Model Zoo, iii) download a short video clip of a dog, and iv) run inference while displaying keypoints. v) remove the temporary folder.
Processor
(if desired)DLCLive
objectfrom dlclive import DLCLive, Processor
dlc_proc = Processor()
dlc_live = DLCLive(<path to exported model directory>, processor=dlc_proc)
dlc_live.init_inference(<your image>)
dlc_live.get_pose(<your image>)
DLCLive
parameters:
path
= string; full path to the exported DLC model directorymodel_type
= string; the type of model to use for inference. Types include:
base
= the base DeepLabCut modeltensorrt
= apply tensor-rt optimizations to modeltflite
= use tensorflow lite inference (in progress...)cropping
= list of int, optional; cropping parameters in pixel number: [x1, x2, y1, y2]dynamic
= tuple, optional; defines parameters for dynamic cropping of images
index 0
= use dynamic cropping, boolindex 1
= detection threshold, floatindex 2
= margin (in pixels) around identified points, intresize
= float, optional; factor by which to resize image (resize=0.5 downsizes both width and height of image by half). Can be used to downsize large images for faster inferenceprocessor
= dlc pose processor object, optionaldisplay
= bool, optional; display processed image with DeepLabCut points? Can be used to troubleshoot cropping and resizing parameters, but is very slowDLCLive
inputs:
<path to exported model directory>
= path to the folder that has the .pb
files that you acquire after running deeplabcut.export_model
<your image>
= is a numpy array of each frameDeepLabCut-live offers some analysis tools that allow users to peform the following operations on videos, from python or from the command line:
resize
or pixels
parameter. Using the pixels
parameter will resize images to the desired number of pixels
, without changing the aspect ratio. Results will be saved (along with system info) to a pickle file if you specify an output directory.dlclive.benchmark_videos('/path/to/exported/model', ['/path/to/video1', '/path/to/video2'], output='/path/to/output', resize=[1.0, 0.75, '0.5'])
dlc-live-benchmark /path/to/exported/model /path/to/video1 /path/to/video2 -o /path/to/output -r 1.0 0.75 0.5
dlclive.benchmark_videos('/path/to/exported/model', '/path/to/video', resize=0.5, display=True, pcutoff=0.5, display_radius=4, cmap='bmy')
dlc-live-benchmark /path/to/exported/model /path/to/video -r 0.5 --display --pcutoff 0.5 --display-radius 4 --cmap bmy
deeplabcut.benchmark_videos
and deeplabcut.create_labeled_video
(note, this is slow and only for testing purposes).dlclive.benchmark_videos('/path/to/exported/model', '/path/to/video', resize=[1.0, 0.75, 0.5], pcutoff=0.5, display_radius=4, cmap='bmy', save_poses=True, save_video=True)
dlc-live-benchmark /path/to/exported/model /path/to/video -r 0.5 --pcutoff 0.5 --display-radius 4 --cmap bmy --save-poses --save-video
This project is licensed under the GNU AGPLv3. Note that the software is provided "as is", without warranty of any kind, express or implied. If you use the code or data, we ask that you please cite us! This software is available for licensing via the EPFL Technology Transfer Office (https://tto.epfl.ch/, [email protected]).
This is an actively developed package and we welcome community development and involvement.
If you want to contribute to the code, please read our guide here, which is provided at the main repository of DeepLabCut.
We are a community partner on the . Please post help and support questions on the forum with the tag DeepLabCut. Check out their mission statement Scientific Community Image Forum: A discussion forum for scientific image software.
If you encounter a previously unreported bug/code issue, please post here (we encourage you to search issues first): https://github.com/DeepLabCut/DeepLabCut-live/issues
If you utilize our tool, please cite Kane et al, eLife 2020. The preprint is available here: https://www.biorxiv.org/content/10.1101/2020.08.04.236422v2
@Article{Kane2020dlclive,
author = {Kane, Gary and Lopes, Gonçalo and Sanders, Jonny and Mathis, Alexander and Mathis, Mackenzie},
title = {Real-time, low-latency closed-loop feedback using markerless posture tracking},
journal = {eLife},
year = {2020},
}