search

craymichael / CNN_LCD

CNNs for Loop-Closure Detection on the Oxford New College and City Centre Datasets
GNU General Public License v3.0
58 stars 14 forks source link

The similarity matrix is all 1? #1

Closed httbupt closed 5 years ago

httbupt commented 6 years ago

run "python cnn_lcd.py overfeat".
Why did I get this? The similarity matrix is all 1. simplot_city_overfeat

craymichael commented 6 years ago

Hi, do you have any other log output from the run? This could arise from a variety of sources (missing dependencies, incorrectly placed files, etc.). Also, how long did the run take to complete? Thanks

httbupt commented 6 years ago

Run time of 100 images in 5 minutes Log file is shown below

1

craymichael commented 6 years ago

Hmm, everything looks alright other than the segmentation fault. After printing the results the only thing that is run is plotting of the PR-curve so I'm unsure why that would cause core to dump. Regarding the similarity matrix, it is possible that the median filtering is the issue; try running with the argument --sweep_median so that other filter sizes are used. Otherwise, this might indicate the Overfeat Python API could be the culprit; expected results from the alternative TensorFlow models in Python 3 with slim.model would confirm such.

httbupt commented 6 years ago

Using TensorFlow model in python 3，the accuracy of experimental results is very low。 precision-recall_curve_inception_v4

craymichael commented 6 years ago

Yes, this is to be expected with these models; I was not able to obtain the same results as the paper on either dataset. As far as I can tell, I have the experiments set up identically and have implemented their method as described, and additionally have added median filtering to clean up results (usually improves AP). Anyway, it appears that the code is working and the problem lies with the Overfeat model. I would attempt to reinstall it using the GPU version if you can. The model Python API is unfortunately quite slow and I only refrained from using the TensorFlow implementation due to a lack of easily loadable trained ImageNet weights.

httbupt commented 6 years ago

Could you let me see the results of the correct experiments you got using overfeat，Three experimental results。

craymichael commented 6 years ago

Hmm, I just pulled the Overfeat code and I appear to be getting the same issue; it appears that the output of every image is the same, which is why the matrix is all 1s. I'll look into it further

craymichael commented 6 years ago

Hi, I opened an issue with sermanet about this. Their implementation worked when I was running experiments, but it does not appear to work anymore (it's possible a dependency update has broken things).

craymichael commented 5 years ago

Closing per source of the issue originating from OverFeat repo.