quantombone / exemplarsvm Goto Github PK

I need to write an explanation of this...

Since LR flips are now inside the main code execution loop and NMS is done right after localization, localizemeHOG will not do NMS across original/LR detections.

Not sure this is a problem, since the final NMS across all exemplar will take care of this.

It seems that something is messed up during calibration and M-estimation, because the final results aren't giving me the boost I was seeing before. I changed a few things for the NIPS submission, and perhaps I introduced a bug.

initialized raw models based on fixedframe "f" framing, or goalsize "g" framing can be one of:

• raw models (normalized HOG)
• NN models (nnmode = 1), which uses cosine angle NN measure
• trained to produced "-svm" models

Thus, there is no need to append NN to the name.. in fact, initialization doesn't know anything about

memex short paper is already done, need to remove milestone

There should one place where global parameters are set, this will help the code live much longer

if we use the most recent hog descriptor code, it will be 32 numbers because there is the 'outside' image bin present.
should we go with it?

This rarely comes up, but better to address now than feel the pain later.

init_params is missing in function signature while calling init_function in exemplar_initialize (line 90).

P.S: Is this right place to report and track bugs?

Do the updates

hi, I download the code, and wanna train the voc2007 classifiers for myself, such as bus, bicycle etc.
my computer is intel i5(2 cores, 4 threads), and memory is 6GB, OS is linux federal.
I have run the voc_demo_esvm for almost 10 hours, but it seems that it is still training.
So, How long do you think my computer should run for just a class, such as bus in VOC 2007 dataset?

for some reason the line drawing is broken in the bus transfers.. I can display them on the exemplar (left pane) just fine, but either the xform or the logic is messed up on the right hand side....!

directories should reflect whether we are applying to in-class images or not. For example, apply_all_exemplars should produce either:
trainval+car_...
or
trainval_...

There should be a way of doing the evaluation without writing/reading files from disk.

This has been requested by multiple people, a simple API to train from your own data source.

To better handle nn files, just append the 'nn' mode to the models_name

For Siggraph

show per-test image results also...

Hi, I have download the code and wanna test the result on VOC2007 . after that I want to use Exemplar SVM in my research work. BUT I do not find the demo code as in the read me file says: For evaluating the PASCAL VOC 2007 pre-trained exemplars, see the notes ...and the main evaluation function in [exemplarsvm/demos/voc_demo_apply.m] .

would you please give me the apply demo code or just tell me how to use the code to detect an object(such as bus) in a image?

Thanks very much!

Once we have the M-rescored boxes, the top one is just a 'prototypical' view.. the raw local associations are better for actually transferring over stuff.

is MAXDIM the culprit? the frames are often too big and I feel they should be smaller

Some functions, such as input streams refer to pascal, but are actually pascal-free now. They should be updated.

Hi,

When trying to run your code, I get an error saying

"Undefined function or variable 'features' "

Error in esvm_detect>esvm_detectdriverBLOCK (line 255)
templates = zeros(S(1),S(2),features,length(models));

"features" looks like an integer value which has not been initialized. Can you tell me what value it should take?

I am running this code on Windows XP.

kill the dupes and gain AP

The class is already in the models name, and thus it shouldn't be mentioned in the model writing functions... Currently there is a duplicate of [cls +'.' cls] here..

the browser will have to be mostly javascript based, since we cannot generate all of the images beforehand

directories should reflect whether we are applying to in-class images or not. For example, apply_all_exemplars should produce either:
trainval+car_...
or
trainval_...

Just create a directory, and place hopefully meaning-ful files in it. Or better yet, you can remove what is inside the directory sketches (old code interface anyways) and place something in it that at least over the sketches.
-T

It seems that I'm only using the stripped file... which is also much smaller

The naming is confusing because the function is called "strip_model" but it actually strips a cell array of models.. This needs to be updated.

These functions are application specific and should be a part of your application, and not hard-coded in the codebase. Our codebase provides bg functions for common datasets such as:
PASCAL VOC: get_pascal_bg
SUNS: get_suns_bg
SUN09: get_sun09_bg
raw_directory: get_directory_bg

NOTE: the files written do not indicate whether we are in NN mode or not
nn mode should have something appended to the end, just as '-nn-cosine or '-nn-normalizedhog'

need to create instructions on how to compile files in the separate projects

I have downloaded the most recent code from the git repository into matlab 2011 added everything to the path and have been unable to run any of the demos. Would it be possible to get some setup documentation?

Thank you,
Bryan

the baseline code should be all set in place

This simple calibration trick doesn't require any held-out data. Use this as a comparison with no-learning.

When there is a large number of exemplars, it is better to use chunk matching as opposed to independently-ran slides...

Look at the old ddip repository, and try the calibration-free M-estimation method.

Need a way to prune the NN files...they are really intractable

james wanted this (and a whole lot of other people).. this will make a killer blog post...

I should make a pdf with some pictures, and a citation to my blog (which will hopefully get people to cite my blog)

calibration results should be saved and cached just like the grids

both the VOC exemplar initialize, and the manual initialize from a sequence are both near-duplicate functions.. they should be merged or internal functions shared...

It is also desirable to have the infrastructure not be dependent on VOC data.

Exemplar initialization should have a dataset-dependent part which creates a sequence of (I,bb,name) triplets, and an algorithm-dependent part which uses a specific exemplar framing algorithm.

they dont look nice in the browser

MAXITER is assigned in the mining_parameters, but it seems that it is not used being used in the training process as an upper cap of iterations. I just saw that even if my upper limit was 10 for iterations, I had mat files from 25-30th iterations.

I even couldn't find a check using MAXITER in train_all_exemplars. Though an an easy fix, but am I missing something?

I've actually come across your paper before, but while searching for orange SVM on github I came across this.

Do you ever have plans to port it to something outside of matlab?

Just to maintain a list of hard-coded params such that we both are on the same page.

MAXDIM is hardcoded in https://github.com/quantombone/exemplarsvm/blob/master/initialize_goalsize_model.m
I prefer a different size for different experiments. That can be moved to init_params.

During mining, when we look at the console output, we see something like #seen=0001/0010 being shown. However, this is only for the current iteration, so when we walk away from the computer and take a look at the console, it is not clear what iteration we are on.

I think a better alternative is to show the TOTAL number of mined images across all iterations. I think because we have the mining_queue (which tells us what images are left) and the train_set (which shows us the total number of images), we should be able to do this easily.

Currently a cell array of matrices is returned, but a matrix is actually more appropriate since I'm calling cat(2,feats{:}) on it all the time...

During mining and testing, we should make sure that fg is used for "foreground" sets and bg for "background" sets. It might be confusing for a new user if they see bg used instead of fg.