TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Cleaning algorithm finds 20% of errors in major image recognition datasets

221 pointsby groarabout 5 years ago

11 comments

CydeWeysabout 5 years ago
Why aren&#x27;t these data sets editable instead of static? Treat them like a collaborative wiki or something (OpenStreetMap being the closest fit) and allow everyone to submit improvements so that all may benefit.<p>I hope the people in this article had a way to contribute back their improvements, and did so.
评论 #22891524 未加载
评论 #22891557 未加载
评论 #22891460 未加载
评论 #22891520 未加载
评论 #22893431 未加载
评论 #22896007 未加载
评论 #22891568 未加载
rathelabout 5 years ago
Nothing is however said about <i>how</i> the errors are detected. Can an ML expert chime in?
评论 #22891152 未加载
评论 #22891119 未加载
评论 #22891214 未加载
kent17about 5 years ago
20% annotation error is huge, especially since those datasets (COCO, VOC) are used for basically every benchmark and state of the art research.
评论 #22891782 未加载
评论 #22892261 未加载
magicalhippoabout 5 years ago
&gt; Create an account on the Deepomatic platform with the voucher code “SPOT ERRORS” to visualize the detected errors.<p>Nice ad.
评论 #22891381 未加载
fwipabout 5 years ago
The title here seems wrong. Suggested change:<p>&quot;Cleaning algorithm finds 20% of errors in major image recognition datasets&quot; -&gt; &quot;Cleaning algorithm finds errors in 20% of annotations in major image recognitions.&quot;<p>We don&#x27;t know if the found errors represent 20%, 90% or 2% of the total errors in the dataset.
评论 #22892090 未加载
评论 #22892083 未加载
kent17about 5 years ago
&gt; We then used the error spotting tool on the Deepomatic platform to detect errors and to correct them.<p>I&#x27;m wondering if those errors are selected on how much they impact the performance?<p>Anyway, this is probably a much better way of gaining accuracy on the cheap than launching 100+ models for hyperparameter tuning.
frenchie4111about 5 years ago
Best I can tell, they are using the ML model to detect the errors. Isn&#x27;t this a bit of an ouroboros? The model will naturally get better, because you are only correcting problems where it was right but the label was wrong.<p>It&#x27;s not necessarily a representation of a better model, but just of a better testing set.
评论 #22891810 未加载
benibelaabout 5 years ago
These things are why I stopped doing computer vision after my master thesis
jontroabout 5 years ago
Weird behaviour on pinch to zoom (macbook). It scrolls instead of zooming and when swiping back nothing happens.<p>Another example of why you should never mess with the defaults unless strictly necessary.
groarabout 5 years ago
Using simple techniques, they found out that popular open source datasets like VOC or COCO contain up to 20% annotation errors in. By manually correcting those errors, they got an average error reduction of 5% for state-of-the-art computer vision models.
评论 #22891128 未加载
m0zgabout 5 years ago
An idea on how this could work: repeatedly re-split the dataset (to cover all of it), and re-train a detector on the splits, then at the end of each training cycle surface validation frames with the highest computed loss (or some other metric more directly derived from bounding boxes, such as the number of high confidence &quot;false&quot; positives which could be instances of under-labeling) at the end of training. That&#x27;s what I do on noisy, non-academic datasets, anyway.