[10:02:29] @AbdealiJK: missing the -dryrun capabilities of bulk_bot ... is there an alternative now? [10:02:56] There is nothing right now - I thought as it's bulk it wouldnt be needed [10:03:15] can add it if you want [10:04:09] May be for beginners during conpherence ... Sometimes it's nice just to disable writing to wiki for testing. No problem, thanks! ;) [10:14:43] alright [10:14:59] DrTrigon, as you see at https://travis-ci.org/pywikibot-catfiles/docker-file-metadata/jobs/145192714 [10:15:11] Python-OpenCV is installed, but the python-opencv tests are giving incorrect results [10:15:19] So, something fishy is going on there [10:16:27] You mention that regarding my comment on the pull request? [10:17:17] AbdealiJK: ^^^ [10:17:25] Yep [10:17:49] DrTrigon, ^ [10:19:09] Ok, perfect - that is a way better solution! As requested...great! :)) [12:39:51] DrTrigon, You had mentioned that we could remove faces which are small if it doesn't occur in groups yesterday [12:40:34] I find there are a lot of images where there is 1 small face. [12:40:36] Like: [12:40:36] https://commons.wikimedia.org/wiki/File:Pays-Bas,_Amsterdam,_Rijksmuseum,_Mercure_et_Psyché_d%27Adriaen_de_Vries_dans_les_jardins_du_Musée_et_(26345322991).jpg [12:40:42] https://commons.wikimedia.org/wiki/File:Peter_Nugent_at_the_Rachel_20th_telescope.jpg [12:44:13] AbdealiJK: What is "a lot"? Do they get detected by landmarks and haarcascade? [12:46:40] It depends, some are detected by dlib only [12:46:45] others by haar only [12:46:47] or both [12:47:05] By "a lot" you had mentioned if only 1 face exists in the image, and the face is small disregard it [12:47:16] i.e. small faces only happen in things like team photos or concerts [12:47:36] DrTrigon, ^ [12:48:24] AbdealiJK: "I find there are a lot of images where there is 1 small face." How many do you find in % per scanned files. [12:48:34] * ? [12:49:21] Ah. [12:49:22] Asked differently: ... [12:49:48] ... if you follow my rule of discarding them; How many % do you loose? [12:50:00] I've seen 9 correct small faces, 1 wrong small faces in ~30 faces that have been detected [12:50:27] What I mean by "small" is less than 5 % of the file size [12:50:32] in terms of area [12:50:38] hmmm [12:51:04] Can we use only the ones that get detected by landmarks AND haarcascade? Would that help? [12:51:28] Checking - 1 moment [12:52:23] Out of the 9 images, only 2 are detected by both [12:52:37] and the wrong one? [12:52:57] The wrong one is a little unusual :P There are 2 faces in that. 1 is by dlib, 1 is by opencv - both are wrong though [12:53:28] are they at the same position? (roughly) [12:53:40] Nope, at different corners [12:53:57] So, I'm looking at logs at -newimages [12:54:01] ok so that could be a possibility to improve at least ... [12:54:17] I can't seem to find many images where the face is small and wrongly detected [12:54:17] ... but given the numbers you mentioned 9 1 30 ... [12:54:32] (exactly, nod) [12:54:48] ... I would say the error rate is acceptable for the moment ... [12:55:02] ... other tricks for improving I can think of ... [12:55:28] ... would include using other cascades too (eyes, mouth, nose, ear) ... [12:55:46] ... or the "wiggeling" and secaling thingy we discussed once. [12:56:03] * scaling * [12:56:09] The other cascades are indeed being used right now [12:56:27] what do they give on the 9 and 1 images [12:56:31] ? [12:57:17] So, the dlib score is 0.05 ish [12:57:47] The 1 error has a dlib score of 0.188 [12:58:12] The haarcascade in the 1 error one doesnt have any eyes, ear, etc [12:58:29] (ok, good) [12:58:29] So, the haarcascade face isnt currently being considered a "valid" face [12:58:55] I've already put a threshold of 0.4 ish for dlib face detection [12:59:05] So, I am letting it be as a outlier and moving on for now [12:59:22] do the correct small faces have eye, etc. haarcascde hits [12:59:41] Yes, they do [12:59:48] so here you have it! [12:59:54] :))) [13:00:24] The dlib score of 0.188 would still add the wrong small face into Category:Human faces though [13:00:45] so ignore dlib for small faces? [13:01:00] But that would remove about 4 of the correct small faces :/ [13:01:45] So hence I'm letting it be [13:01:46] 4 out of 9 - you still get more than 5'% [13:01:50] 50% [13:02:10] I'd rather have 1 error out of 30 than reduce it to 25 out of 30 [13:02:29] Know what we can do? [13:02:35] hm? [13:02:59] That is a perfect situation for human-assisted vs. automatic, what do you think? [13:03:07] yep [13:03:37] cool! so we have already settings for that. that is +1 on the income side! ;)) [13:03:55] and I agree; 1 wrong out of 30 is around 3% wrong... [13:04:02] ...that is good for CV! [13:05:23] DrTrigon, so it seems that opencv has some issue in the dockerfiles [13:05:43] WHen free, could you debug ? [13:05:46] As I can't [13:05:59] the strange one you mentioned ... somehow I expected that at some point ... [13:06:03] As I can't load them on my computer ... it's a little difficult for me to do right now [13:06:35] * DrTrigon away for 5 mins [13:14:44] AbdealiJK: Sure I'll have a look! Is it urgent? [13:14:53] Nope, not urgent [13:15:04] perfect - I'll do it! [13:15:20] AbdealiJK: Is that ok for know, anything else? (Are you happy? ;) [13:15:26] I'm good :D [13:15:38] Niicce! [13:15:49] see ya then!