eths / source /studio /find_photos.py
Ethscriptions's picture
Daily English Live Photos: 2 sets x 9 live photos (static site)
8ea8b32 verified
Raw History Blame Contribute Delete
3.49 kB
"""Find real CC-licensed photos (Openverse: Wikimedia Commons / Flickr) and rank them with CLIP."""
import io, json, os, sys, time, urllib.parse, urllib.request
import numpy as np
from PIL import Image
sys.path.insert(0, '/workspaces/studio')
import clip_score as C
ITEMS = json.load(open('/workspaces/studio/items2.json'))
OUT = '/tmp/work/real'
os.makedirs(OUT, exist_ok=True)
UA = {'User-Agent': 'LiveWordsStudio/1.0 (educational live photo maker)'}
NEG = ['a drawing or illustration', 'a diagram or logo', 'text and a document', 'a product catalogue cutout on white background',
'a black and white historical photo', 'a museum exhibit']
def get(url, timeout=60):
return urllib.request.urlopen(urllib.request.Request(url, headers=UA), timeout=timeout).read()
def search(q):
res = []
for page in (1, 2):
p = urllib.parse.urlencode({'q': q, 'license_type': 'commercial', 'category': 'photograph', 'page_size': 20, 'page': page})
try:
d = json.loads(get('https://api.openverse.org/v1/images/?' + p))
except Exception as e:
print('search err', q, e); break
res += d.get('results', [])
time.sleep(1.5)
return res
report = {}
for it in ITEMS:
if len(sys.argv) > 1 and it['id'] not in sys.argv[1].split(','):
continue
seen, cands = set(), []
for q in it['q']:
for r in search(q):
if r['id'] in seen or not r.get('width') or not r.get('height'):
continue
seen.add(r['id'])
w, h = r['width'], r['height']
cw = min(w, h * 3 / 4) # width of the 3:4 crop
if cw < 900:
continue
cands.append(r)
pos = C.text_emb([it['clip'], 'a candid everyday photo of ' + it['clip'][11:]])
neg = C.text_emb(NEG)
others = C.text_emb(['a photo of ' + o['en'] for o in ITEMS if o['id'] != it['id']])
scored = []
for r in cands[:45]:
try:
th = Image.open(io.BytesIO(get(f"https://api.openverse.org/v1/images/{r['id']}/thumb/", 40)))
except Exception as e:
continue
# score the centre 3:4 crop, which is what ends up in the frame
tw, tht = th.size
cw = min(tw, tht * 3 / 4); chh = cw * 4 / 3
th = th.crop(((tw - cw) / 2, (tht - chh) / 2, (tw + cw) / 2, (tht + chh) / 2))
e = C.image_emb(th)
sp = float((pos @ e).max() * 100)
sn = float((neg @ e).max() * 100)
so = float((others @ e).max() * 100)
portrait = r['height'] / r['width']
score = sp - 0.5 * max(0, sn - 20) - 0.5 * max(0, so - sp + 2) + (1.0 if portrait >= 1.2 else 0)
scored.append(dict(id=r['id'], title=r['title'], creator=r.get('creator'), license=r['license'],
license_version=r.get('license_version'), license_url=r.get('license_url'),
landing=r.get('foreign_landing_url'), url=r['url'], w=r['width'], h=r['height'],
source=r.get('source'), sp=round(sp, 1), sn=round(sn, 1), so=round(so, 1), score=round(score, 2)))
scored.sort(key=lambda x: -x['score'])
report[it['id']] = scored[:6]
print(it['id'], it['en'], len(cands), 'cands')
for s in scored[:4]:
print(' ', s['score'], s['sp'], s['sn'], s['so'], s['license'], s['w'], s['h'], s['source'], (s['title'] or '')[:60])
json.dump(report, open(OUT + '/candidates_%s.json' % it['id'], 'w'), ensure_ascii=False, indent=1)