prbonn / online_place_recognition Goto Github PK

I extract Alderley datasets of Sunny_Day and Rainy_Night.
Firstly, I compared two identical sequences, so the expected result should be a diagonal line, which it is. So this is correct. However, the two windows have no photos been retrieved, which said "Pixmap is a null pixmap".

I set Sunny_Day as reference and RainyNight query. However, I get a vertical line in matching matrix, which means that the car is staying within the query sequence. But in case of Alderley dataset it is not the case.
Olga gave me two situations that might be happening here:
"

1. Your non-matching cost is too high. This means that you allow images with "high" matching cost to be considered as a match and the system do not differentiate between good and bad matches. To check this try to gradually reduce the non-matching cost until you see blue nodes appearing on the line and check if the shape of the line is changing. If, however, only small amount of red nodes are left, than you restrict your search to only peak very good matches.

2. Something is wrong with the way the costs are computed. The only way this search algorithm can give reasonable results is if the "underlying" cost matrix actually has a pattern ( this you in theory should not compute for online case). To check if this is the case you can construct the full matching matrix and see if you as a user can see a pattern in the matrix. For this you can use apps/create_cost_matrix code. If not, something is wrong with cost computation or the features + metric is not good enough for dealing with such challenges (but I don't think this is the case for Alderley and OverFeat (10th layer)).
   "
   I am trying these two and will update later.
   Thanks so much to Olga.

When I create cost matrix with "./create_cost_matrix config.yaml" in folder online_place_recognition/build/bin/, it said:
"
dyld: Library not loaded: @rpath/libopenblasp-r0.2.19.dylib
Referenced from: /Users/zhanghui/anaconda/lib/libopencv_stitching.3.2.dylib
Reason: image not found
Abort trap: 6
"
Besides, my config.yaml writes as below:
path2ref: reference/OverFeat_features/
path2qu: query/OverFeat_features/
querySize: 999
expansionRate: 0.1
fanOut: 1
nonMatchCost: 2.0
path2refImg: reference/images/
path2quImg: query/images/

Hi, Thanks for the code.
We want to implement a off-line place recognition. That is, the reference is first gone through to generate a database, then the query will be gone through to achieve place recognition. In this case, the system is more like a loop detection framework. Can u tell us what should we do?