ICLR14: P Sermanet: OverFeat: Integrated Recognition, Localization and Detection using ConvNets

What this episode is about

The presentation focuses on the OverFeat system, which integrates recognition, localization, and detection tasks using convolutional networks (ConvNets) for image processing. Key advancements include the use of enhanced feature extraction, multi-scale training strategies, and competitive performance comparisons that highlight the system's effectiveness in achieving state-of-the-art results in visual recognition tasks.

Key takeaways

• In deep learning, size matters; more connections often lead to state-of-the-art features and results.
• Surprisingly, just 50 minutes tweaking interfaces can transform performance in object recognition — proof that effort counts!
• Localization is the unsung hero of AI: crucial but often overshadowed by more glamorous classification tasks.
• Forget perfect squares! Multiscale approaches redefine object detection, catching tiny players in a big picture.
• Combining quick generic detection with detailed models streamlines performance, proving that speed and accuracy can coexist.

Who speaks in this episode?

• P. Sermanet (Host) · 14 min speaking

What is discussed, and when

1. 3:00 – 6:00: Tiny Network, NC, UVA, Amazon
2. 6:00 – 9:00: Bing box
3. 9:00 – 12:00: Bing box
4. 12:00 – 15:00: Aion

Notable quotes from "ICLR14: P Sermanet: OverFeat: Integrated Recognition, Localization and Detection using ConvNets"

• 50 minutes of cing to interface feet and 20 bucks in Amazon and what so bottom line is to put these guys from now deep learning has to all right now about location.

  about Amazon · at 3:33

• Tiny control here and you will be penalized for so the results for loation we won the loation 30% error and detection we didn't win during the challenge butek later we get 23% me aage proced and the UVA.

  about NC · at 4:30

• Recute everything so for example this Tiny Network here if you were to add some extra pixels you only have to compute the area if you already did that part so you know this is why it's okay to it's design that detection the way.

  about Tiny Network · at 5:31

• Evidence for objects so when you slide comments if for each window as you slide you produce Bing box and you can acculate Bing boxes and be really sure what the answer and I'm predicting B boxes is really similar to the work that did in.

  about Bing box · at 7:44

• Aion and I think just by swapping inets anding all methods I think if we keep this going up both in terms of mods and data we can keep same.

  about Aion · at 13:40

Who and what is mentioned in this episode?

Companies & orgs

• NC — Company · Referenced organization · 1 mention · first at 4:51
  Mentioned alongside UVA in the context of object detection research.
• UVA — Company · Referenced organization · 1 mention · first at 4:51
  Mentioned in the context of object detection research and competition results.
  Tiny control here and you will be penalized for so the results for loation we won the loation 30% error and detection we didn't win during the challenge butek later we get 23% me aage proced and the UVA.
• Amazon — Company · Incidental mention · 1 mention · first at 3:36
  Mentioned in passing as a platform where services or products can be purchased.

Products & tech

• Tiny Network — Product · Research tool · 1 mention · first at 5:33
  Referenced as a small-scale network used in object detection research.

Concepts & topics

• Bing box — Concept · Technical concept · 2 mentions · first at 7:51
  Discussed as a method for accumulating evidence for object detection in images.
• Aion — Concept · Technical concept · 1 mention · first at 13:37
  Mentioned as a term used to describe advancements in object detection networks.

Full transcript with timestamps

0:01 um so I will start by saying that is the name of the system that's also the name of the of the Shelf F extractor that we released um which is uh this pict is to be for vision people want to train their own network um so we team to enter all three tasks for classification loation and and one of the point of this word is uh it's a nice example of transer running where we started with features done fromation and we actually managed to learn state of the art um regress on these features to actually predict coordinates of boxes so these are the results um for classification top results up on Matt's Min here andw Howard here and so OB the black box is PR competitive and it's based on the mod by Alex um and it's really similar I mean the one of the main difference is just getting up so we had more features more a small stri for the first and we went up to 1.5 million 5on connections um and another improvements

1:32 we did is multiscale W instead Al 10 New multiv View and this is a more um principal way of doing U multiv View and I explain later that you know to conet you can actually do dance it's Prett um so in the past six months there's been a tremendous amount of work uh get using comments um and I put this side together because this group from stock they use the word sing to describe next uh so I have three slides of results for the six months uh so what they did here they took the box and they put on top and they kept it pretty simple on purpose uh just to see how I will do on other task so this they did recognition recognition detection and they got uh competive or stere results um and it was really surprising to them and um I guess you know the lesson is that Futures are really um so good work byor people here we are uh and two weeks ago this guy w Galaxy and so also my own challenge was the

3:13 drugs C have using Fe and what's interesting is that the top performers they most of them use Fe or a combination of um and more interestingly second best results uh this guy he said he just thre 50 minutes of cing to interface feet and 20 bucks in Amazon and what um so bottom line is to put these guys from from now deep learning has to all right now about location detection uh so first of all what is calization um so it's essentially a hor verion of the classification Tas because you have in addition to joining the right Tass you have to give a type box around the object um but you you also uh pretty much guaranteed you have at least one object and it's going to be big image uh so it's not hor as detection uh where you can have any number of objects and they can beze tiny control here um and you you will be penalized for uh so the results for loation we we won the loation 30% error uh and detection we didn't win

4:47 during the challenge butek later we get 23% me aage proced um and the UVA and NC teams have different which other um I just want to make the point thats are good for detection because um they're basically contion all the way uh including the which you can see as one by one and uh which means that you can have any uh input size of your images and you don't you can share the results the years and you don't need to recute everything so for example this uh Tiny Network here if you were to add some extra um pixels you only have to compute the area if you already did that part so you know this is why it's okay to it's design um that detection the way uh traditional object detection goes uh you do sing window uh then no maximum supression you only keep one uh the strongest activation and offline W scrapping so offline W scrapping means um you're going to do a pass over your all your images and you find the worst

6:12 answers uh and then you're going to uh so some of problem the problems with this is that you get um fixed aspect pressure windows so like these squares here so this is a multiscale uh dense and um and this approach is also really sensitive to PO um and offline bopping is very complicated uh and if you use Nonet models then you're going to have to rec to home model every loti um so to upate some of these parts recent um work uh has been doing this objectness uh measure where SC using is what I where you um do a quick and cheap uh generic detection first you find interesting candidates and then you want more intense model on on um so you can do it with segmentation Nets uh and this speed things up uh it reduces theity quite a lot and um and so this this was your approach by U the second best um we didn't in um so this is uh I started doing a b box prediction this for pedestrian and the idea was more was less about speing

7:41 things up was more about uh increasing the evidence um for objects so when you slide comments uh if for each window as you slide you you produce Bing box and you can acculate Bing boxes and be really sure what the answer um and I'm predicting um B boxes is really similar to the work that did in 2004 where they were um trying to regress to this face manifold U and this manifold is trying to predict the presence and the grow of Faces in uh so it's pretty much the same um and yeah by the way these people are young sisters um it's my grandmother grandmother um so so the way our works um this is M scale umet studing window and each of these boxes are one and so we we have this um feature exing phase which is common for both um classication and opiz so this the all the features are shared and then you have two different uh outputs so you have a outut for tation so for each of these box you're going to say is it a

9:03 bear or is it a whale so yellow is there and P is well um and then the second output is going to predict project uh bing box and that's get um so now the next thing is uh you know if all the b bs are all aligned then you're pretty sure that uh this should be the rights and if they all have the same color then you're pretty sure it's the right test um if if it gets know all messy then maybe they don't agree and it's it's more like a false this is a very good way to separate true um this is yeah one form of voting for um and this shows so the better it's align the the more sure we are and in this example it shows that you know it still works for when you have multiple objects you when you in C and you start your window um the boxes are going to you know distribute themselves natur each object so it does work for multiple objects um and then you after you have you have all these windows you can merge

10:27 them based on the how the line uh different measures and you end up in that case you end up with this one really confidence box that this is a there at this location um so one thing I want to mention is that when you do you if you have lot of something you might end up with a big gap and it's detction so what you want to do is increase uh decrease the size strike uh and there's a trick for that um that that system uh so for good strapping so we don't do this offline nothing because uh it's uh when the get B it's get comp um and doing online is better because you that constantly doesn't report any storage um uh so you could you know for you could say for that image I'm going to backrop all these boxes um but you might not want to do because uh that might bu too much for each image so instead what maybe I should do is back I can do worse false positive in that image woral negative and oneand negative and one

11:44 posi um okay some examples of detection so so sometimes we get duplicates it's not perfect yet and of course it also it can works the you know some changes Cas table um okay this one's pretty U interesting I think because we barely see the instrument and yet we are to CL and um this one is also good because um uh there's a quite a bit of overlap between the two birds and yet we use two different outs so the right is a ground and left is our um so it does work well over same have some over dos um and the next thing is that most of the time the files make a lot of sense so this the gra is snake the sake gr but it does you know look like it um and we a remote control from the fridge and it kind looks like control um but yeah we don't find and tables not um so IM is pretty hard uh there's know dos labels and some pictures um like there's a mic here person drum and yeah in this case we only find person so this it's really

13:20 interesting data set uh for example yeah same for fo umies we found like three bananas and one hot do and one times more so uh there still some way to go uh so to recap the has become Aion and I think uh just by swapping inets anding all methods um I think if we keep this going up uh both in terms of mods and data we can keep same um so we show that we can do nice learning you know learn different tasks you know cation from futes train for classification um which should good protection uh accumulating uh prediction is increasing increasing results um and doing Dan adding Windows is important to get good results so which is why I think the next step might be learning attention uh we get if we know exactly what to look then we can really focus on on exact right boxes and that we not only increase the speed but also increase the access and that's it thank you you predict the window size at which

14:53 location uh it's foures so Center and and scale Al so it's not necessarily centered on the pixel response to the um yeah I mean there's different things you can do and they all can work so do XY and I so you mentioned uh the next step is attention I think that's a great idea uh but from your results it also seems like you could benefit from treating it like that structured up where you take advantage of the context of perhaps the other objects that are detected detection of every object that another Direction yes that's something we wanted to do we didn't have time to do uh obviously um uh sometimes you know um like the tennis ball yeah his like we should be able to say you know this yeah but it's it's more it's research uh yes uh use I do understand when you talk about using the positives and negatives for backdrop uh why're not using all the positives uh the ground through posi uh no I mean boot stting is more

16:16 about exploring your own mistakes so you infer an image you look at the answers and uh oh the false positive false positive yeah uh but also you still want to train on the posi and but but in that case the false negatives should all be there um was well sometimes you uh right if there was a window you should true posi false negative but the point I want to make is it's not good to train on easy ones much and you want to have good bance and you want find B if there's not other questions let's thank again

Questions about this episode

Who speaks in "ICLR14: P Sermanet: OverFeat: Integrated Recognition, Localization and Detection using ConvNets"?

P. Sermanet (Host).

What is "ICLR14: P Sermanet: OverFeat: Integrated Recognition, Localization and Detection using ConvNets" about?

The presentation focuses on the OverFeat system, which integrates recognition, localization, and detection tasks using convolutional networks (ConvNets) for image processing. Key advancements include the use of enhanced feature extraction, multi-scale training strategies, and competitive performance comparisons that highlight the system's effectiveness in achieving state-of-the-art results in…

What is mentioned in this ICLR episode?

6 people, works and ideas are mentioned, including Bing box, Tiny Network, Aion, NC, UVA, Amazon. Each one on this page links to its own page with every episode that mentions it.

How long is "ICLR14: P Sermanet: OverFeat: Integrated Recognition, Localization and Detection using ConvNets"?

The episode runs 17 min; the transcript is about 2,148 words in 12 timestamped sections.

More from ICLR

• All analyzed episodes
• People mentioned on ICLR
• Topics mentioned on ICLR