Problem with direction sensitive image-data - tensorflow

I´m new here, so please be kind and teach me if I did not provide all the information you need :)
I need to detect objects by object detection. Everything works fine for most of my object classes, but there is a problem:
I have some objects that are sensitive to their direction on the image. More precisely: I have objects that are aligned as "to the right" or "to the left" and should be recognized as such.
I already know that TensorFlow reflects the images randomly - which destroys my alignment, because left is right and right left. I come to this conclusion because all other objects are detected in a good amount and quality.
Therefore I want to "forbid" TensorFlow to mirror the images and only to rotate them instead.
I have already searched the Python classes and identified the function "tf.image.flip_left_right" which does the mirroring.
Now I don't know which of the many many files and places, where this method is called, I have to change (none of the scripts are mine, they all come natively with TensorFlow!)
Here are the scripts the method was called:
...\models\research\object_detection\core\preprocessor.py (1 hit)
...\models\research\object_detection\core\__pycache__\preprocessor.cpython-36.pyc (1 hit)
...\models\research\object_detection\training\events.out.tfevents.1559030641.R233689 (60 hits)
...\models\research\object_detection\training\graph.pbtxt (24 hits)
...above that: 36 Hits in every model.ckpt-XXXXX.meta-file
(which I dont think they are important for this problem?)
Has anyone had any experience with it and can give me any advice?
EDIT: I´m using TF in version 1.13

You should have a .config file somewhere which defines the augmentation applied to your images. See for example this file.
Just remove random_horizontal_flip {} and you are done.

Related

How to sculpt with odd linear artefacts

I am currently learning how to sculpt in Blender; working on my own projects after completing BlenderGuru's Beginner & Intermediate classes, and some of Grant Abbitts videos with pleasing results. I am trying to sculpt a plasmapistol with a skull on it, which can be seen in the reference photo that I have provided.
However, when I sculpt, I get these really odd linear artefacts (See picture below, circled in black). I added a Subsurf Modifier to the primitive UV Sphere, with the Viewport and Render Values set to 4, so it is a fairly fine mesh. However, these still artefacts occur.
I assume it is due to the stretching of the polygons when I grab the sphere with the Snake Hook tool and deform it to encompass the frontal part of the skull.
EDIT: Whilst writing this comment I went back, and switched on Dynamic Topology with Relative Detail selected.
It appears that I am no longer getting the issues that I was getting last night with the linear artefacts.
Can I confirm that these problems are a result of having the incorrect Dynamic Topology settings for using the Snake Hook Tool; I was using Constant Detail instead of Relative Detail, or is this being caused by another issue?
Also, any advice on avoiding common pitfalls when choosing the settings in sculpting would be most appreciated.
I will continue to ask this question incase anyone has a similar problem and it can be resolved by reading this.
Sculpt, showing lineations
Experimenting with Dynamic Topology
In Object Mode, does the object have uniform X, Y, & Z scaling? If not, you can apply the scale from the object menu.
Object ‣ Apply ‣ Scale / Rotation & Scale

Does ZedGraph offer any kind of Level-of-Detail Culling behavior?

I've searched and can't find an answer to this question. I could write the code myself to do it, but I don't want to reinvent the wheel. :)
Since ZedGraph uses an IPointList and its indexer for internal data access, you can assign any kind of data structure to it and dynamically change the data that ZedGraph receives when it calls the indexer.
It's a smart architecture, and naturally, it would be feasible to implement a Level-of-Detail system using a custom IPointList where the number of points is culled based on the xScale and yScale of the GraphPane.
This way you can have millions of points loaded, but when the zoomlevel of the graph would show all the points, they can be culled so that ZedGraph is only drawing a few thousand. As the zoom magnification is increased, fewer points would be culled in the region of interest.
I wanted to know if ZedGraph already offers anything like this out of the box. I haven't seen any indication of support for it.
Does anyone know?
I posted about this on Sourceforge and got no response there either.
Then I posted on a fork on Github and got a response. It's here:
https://github.com/ZedGraph/ZedGraph/issues/13
The answer:
There is a naive algorithm that filters points by simply skipping them blindly to reach a target display number.
Of course this naive approach can give completely wrong impressions of what the data looks like when peaks and valleys get dropped in a line graph, for instance. IMHO, an algorithm like that is completely unuseable.
So basically, there is no acceptable built-in culling in ZedGraph at the present time.

Given a pair of images, how to automatically create an animation sequence morphing one image into the other?

Is there a programmatic way to convert two images into an animation sequence (e.g., an animated GIF) like the following example?
This image sequence, taken from a http://memrise.com course, doesn't seem to have manually-edited frames, but seems automatically transformed using some kind shape morphing algorithm. Is there a common term used to describe such an animation or algorithm? Is there a feature in ImageMagick or Photoshop/Gimp that generates such animations, given a pair of images?
Ideally the technique could be scriptable so I could create animations for several pairs of start-end images.
Edit: I have just been told about Gimp's tool under Filters->Animation->Blend, which appears to do the same thing as jQuery morph: each frame i is start + (finish - start)/N*i. In other words, you're transitioning each pixel independently from the start value to the finish value, without any shape morphing. The example gives is more complicated, as it modifies the contours of both images to achieve its compelling effect.
Other examples:
http://static.memrise.com/uploads/mems/32000121024054535.gif
http://static.memrise.com/uploads/mems/225428000121109232837.gif
I have written a tool that doesn't require setting manual keypoints and is not restricted to a domain (like faces). Anyway, the images have to be similar (e.g. two faces or two cars from the same perspective).
https://github.com/kallaballa/Poppy
There is also a web-version created with emscripten.
I generated the above animation using following command line:
poppy flame.png glyph.png flame.png
Although this is an old question, since ImageMagick is mentioned, for anyone who comes here from google it may be worth looking at this imagemagick plugin called shapemorph.
GIMP can't do that directly, but over the years a series of (now poorly maintaind) plug-ins to do that where released by third parties. The keyword for searching for this is "morph" - you should find a bunch of stand alone programs to do that as well, from "gratis" to full fledged Free Software, such as xmorph
Given pairs of vector files (.wmf extension) it is possible to use linear interpolation of shapenodes in Visual Basic for Applications to create frames for GIF animations , though this would take along time to explain. For some examples see
http://www.giless.co.uk/animatorMorphGIFs.htm (it is like a slideshow)
I have made some improvements since then, as well!

Huge images in iOS app without CATiledLayer?

I have an image at about 7000x6000px. I need this to be in a scrollview/imageView in my app, however this is way to huge for display. It is supposed to be a kind of map. I was hoping to keep the size of the app to the minimum, and the image is just about 13mb in .jpg. In .png it is over 100mb, which is unacceptable. Many have suggested CATiledLayer as an option, but I believe this would result in even bigger file sizes. Anyway, I tried to do it with CATiledLayer, and create my own tiles in TileCutter, (tiles in .jpg), and the size wasn't too bad. But I am having errors all over the place. The iOS version of CATiledLayer is a mystery to me, and I can't find a way to solve this. I get an error saying something about the java-equivalent "index out of bounds of array", even though the array has content at that specific index..
It has a method which returns an array. The array contains data of a .plist. Before the return I log out the content of the array, giving me good data. The call is trying to access
[array objectAtIndex:0]
and put it in a dictionary, but throws OutOfBounds. When logged out the whole array, I can clearly see the content, but when logged out
NSLog("%#",[method objectAtIndex]); I get the same exception.
Anyway, CATiledLayer has given me nothing but problems. I have been reverse-engineering the PhotoScroller project with no luck. Anyone have any other solutions?
Thanks.
Apple has this really neat project, PhotoScroller, that uses CATiledLayer and lets you scroll through several images and zoom them. This seemed really neat until I found out that Apple "cheated" and pre-tiled the images (about 800 tiles saved as file in the bundle!)
I had need for a similar capability, but had to download the images from the network. Thus came about PhotoScrollerNetwork. With the TiledImageBuilder you can download (or read from disk) massive images - I even tested a 18000x18000 image - and it works.
What this class does is start tiling the image as it downloads (when using libjpegturbo) or can save the file then tile (takes longer). The class figures out how many levels of detail are needed to show the image at full resolution and sized to fit in the containing view (a scrollview).
The class uses the disk cache to hold the tiles, but uses and old Unix trick of creating a file, opening it, then unlinking it so that the tiles never really get saved - once the class is dealloced (and the file descriptor closed) the tiles are freed (or if your app crashes they get freed too.
Someone had problems on an iPad 1 with its quite limited memory, so the class now throttles its use of the file system when concurrently loading multiple images. I had a discussion with the iOS kernel manager at WWDC this year, and after explaining the throttling technique to him, he said the algorithm (on managing the amount of disk cache usage) was probably the best technique that could be used.
I think those who suggested CATiledLayer are right. You should really use it! If you need a sample project that displays a huge bitmap using that technology, look here: http://www.cimgf.com/2011/03/01/subduing-catiledlayer/
Many technologies we use as Cocoa/Cocoa Touch developers stand
untouched by the faint of heart because often we simply don’t
understand them and employing them can seem a daunting task. One of
those technologies is found in Core Animation and is referred to as
the CATiledLayer. It seems like a magical sort of technology because
so much of its implementation is a bit of a black box and this fact
contributes to it being misunderstood. CATiledLayer simply provides a
way to draw very large images without incurring a severe memory hit.
This is important no matter where you’re deploying, but it especially
matters on iOS devices as memory is precious and when the OS tells you
to free up memory, you better be able to do so or your app will be
brought down. This blog post is intended to demonstrate that
CATiledLayer works as advertised and implementing it is not as hard as
it may have once seemed.

Resizable image resource with embedded cap insets

This is by far not a showstopper problem just something I've been curious about for some time.
There is this well-known -[UIImage resizableImageWithCapInsets:] API for creating resizable images, which comes really handy when texturing variable size buttons and frames, especially on the retina iPad and especially if you have lots of those and you want to avoid bloating the app bundle with image resources.
The cap insets are typically constant for a given image, no matter what size we want to stretch it to. We can also put that this way: the cap insets are characteristic for a given image. So here is the thing: if they logically belong to the image, why don't we store them together with the image (as some kind of metadata), instead of having to specify them everywhere where we got to create a new instance?
In the daily practice, this could have serious benefits, mainly by means of eliminating the possibility of human error in the process. If the designer who creates the images could embed the appropriate cap values upon exporting in the image file itself then the developers would no longer have to write magic numbers in the code and maintain them updated each time the image changes. The resizableImage API could read and apply the caps automatically. Heck, even a category on UIImage would make do.
Thus my question is: is there any reliable way of embedding metadata in images?
I'd like to emphasize these two words:
reliable: I have already seen some entries on the optional PNG chunks but I'm afraid those are wiped out of existence once the iOS PNG optimizer kicks in. Or is there a way to prevent that? (along with letting the optimizer do its job)
embedding: I have thought of including the metadata in the filename similarly to what Apple does, i.e. "#2x", "~ipad" etc. but having kilometer-long names like "image-20.0-20.0-40.0-20.0#2x.png" just doesn't seem to be the right way.
Can anyone come up with smart solution to this?
Android has a filetype called nine-patch that is basically the pieces of the image and metadata to construct it. Perhaps a class could be made to replicate it. http://developer.android.com/reference/android/graphics/NinePatch.html