Just how effective is the search function on Numista?

I was just looking through some the questions on the Exonumia forum and came across this…

https://en.numista.com/forum/topic153990.html 

Having seen on another post that there is now an ‘Image Finder’ on Numista, I decided to find it and try it out.  This is what I found…

Copied and pasted:  

Results:   2 x pages of 110 random coins with only 2 x coins with swords on them (clearly, the swords on the image above are one of the key features).  Only 1 x coin had any Arabic writing on it (another key feature).  Most coins were from China and bore no resemblance to this coin, apart from that some of them had the Chinese flag's wreath on them.  41 x coins were gold (???) and there were a few more that were bronze.  4 x coins had a crown, but only 1 x coin had a shield.  There was even an Elongated Coin and 2 x banknotes.  The image above was obviously not included in the 110 coins listed as it does not appear to be in the catalogue yet.

I then tried to search for the image above along with the words ‘Swords  and ’Arabic', but got nothing (literally ‘’''NOTHING'''').  So, I took out the image and just searched for Swords and Arabic.  This produced 41 x coins, 2 x of which had no Arabic or Swords on them at all.  

I then tried ‘Swords’ alone and got 1150 x coins on 23 pages.

Maybe it was because I tried to do this early on a Tuesday morning, which clearly isn't the best time of day, or day of the week, for the Image finder😁:).   Or, maybe some more work is required on this function.  

So, why not try it yourself?

LDC

One of the problems with the Numista image search is that returns hits with very low probability of being correct. It's more of an “I can't find anything so let's give them everything and let the human decide" approach.

I highlighted another problem with it recently here: https://en.numista.com/forum/topic153609.html#p1205856  where the same picture slightly rotated from the correct position returns nothing useful.

rsirian1

One of the problems with the Numista image search is that returns hits with very low probability of being correct. It's more of an “I can't find anything so let's give them everything and let the human decide" approach.

 

I highlighted another problem with it recently here: https://en.numista.com/forum/topic153609.html#p1205856  where the same picture slightly rotated from the correct position returns nothing useful.

Agreed, it returns a fixed list of results, instead of admitting it does not find good results; when there are results it is very performing (I used it a lot recently to create medieval series like https://en.numista.com/catalogue/series.php?id=8887 and it worked perfectly

The idea to improve would be to display only results beyond a probability threshold

The search by image works differently from what humans would naturally do. Results may not necessarily include swords or Arabic legends, even if they are key features for us.

Indeed, when no results are found, the search will still return the closest matches, even if they are irrelevant.

I tried defining a probability threshold in the past, but it's difficult to choose the right threshold. When a search fails, the probabilities are always low. But some successful searches also have quite low probabilities, yet the correct results are correctly returned toward the top.

Here is an example: The results #1, #3 and #5 correctly identified the city, but with quite low probabilities 23.3%, 22.4% and 21.2%.


To be compared with the OP query, where the top 5 results have probabilities between 25.5% and 23.1%, despite being irrelevant because the correct result is missing in the Numista catalogue.


For easy cases, the probabilities can go above 70%. Here an example with the top 3 results ranging between 71.5% and 69.4%


Therefore I prefer to let humans decide.

rsirian1

I highlighted another problem with it recently here: https://en.numista.com/forum/topic153609.html#p1205856  where the same picture slightly rotated from the correct position returns nothing useful.

That's indeed a big issue. We worked on solving it during the year, but without success so far.