Here's the manuscript on BioRxiv: https://twitter.com/biorxivpreprint/status/1219820808035602433?s=20 … And below a quick summary in tweets!
-
-
Prikaži ovu nit
-
Huge props to
@TrackingActions and@TrackingPlumes for giving us@DeepLabCut! DLC is incredibly powerful and has taken our lab by storm. This project was led by@osturmscience (powered by@Huel) in collaboration with twitterless Lukas "the machine" von Ziegler#ML#DeepLearningPrikaži ovu nit -
Questions: 1) How does the tracking performance of DLC compare to popular commercial systems for rodent behavior analysis? 2) How well do these systems recognize pre-defined rodent behaviors? 3) How easy is it to train DLC to do as well (or better), and how far can we push it?
Prikaži ovu nit -
Approach: We focused on 3 widespread rodent tests (open field, elevated plus maze, forced swim) and generated well-annotated behavior videos (multiple raters). These publically shared videos serve as ground-truth for directly comparing the performance of different systems.pic.twitter.com/BqyfcPHJuS
Prikaži ovu nit -
Of course, you all know how well
#DeepLabCut can track rodents on standard tests like the open field, the elevated plus maze, or the forced swim test. But did you know how poorly some very costly commercial systems perform?pic.twitter.com/XMFveFmzvJ
Prikaži ovu nit -
We then coupled DLC tracking with straightforward post-hoc analysis tools to detect simple behaviors (floating in the FST; head-dips on the EPM). We quantify these behaviors with similar or better accuracy than market-leading commercial systems (our code is freely available).pic.twitter.com/2bTQDTvm2o
Prikaži ovu nit -
We then integrated the tracking data from DLC with supervised machine learning approaches. This allowed us to detect and quantify complex ethological behaviors (like supported vs. unsupported rearing) with human-like accuracy, thus vastly outperforming commercial systems.pic.twitter.com/adkilohyuQ
Prikaži ovu nit -
Our work provides urgently needed benchmarking experiments, comparing commercial systems and DLC to human annotation. The freely available, annotated behavior videos will hopefully serve to push these boundaries. https://zenodo.org/record/3608658 Let us know what you think and please RT!
Prikaži ovu nit
Kraj razgovora
Novi razgovor -
-
-
Have you tried your NN on Ethovision's raw data to classify rearing? The comparison DLC+NN vs Ethovision's Behavior recognition module seems a bit unfair (the module uses a rule-based classifier, I believe).
-
We have not tried. It would be beside the point for us. The unique advantage of DLC is that YOU can decide, which points you track and HOW YOU define behaviors. This is (A) impossible with Ethovision (or other systems), and (B) they charge you mightily for every analysis module!
- Još 2 druga odgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.