cow and pig faces aren’t so easy to tell apart with software originally designed for flatter human faces. and an array of four-legged creatures from china to ireland have been kicking up a fuss when technicians show up to take their pictures to better train the algorithms.for better results, technicians recommend adding side profiles of animals, which have more distinguishing depth and features than those of humans.
developing reliable facial-recognition software depends on getting clear, well-lit and straight-ahead images.
no one told the farm animals.
“there were pigs biting and dragging on my clothes,” says deng changshun, chief executive of a beijing-based developer of animal-recognition software, called innovationai co. “others were attacking me with their snouts.”
cows have tried to hide, licked cameras and untied equipment cables with their tongues. pigs have squirmed out of reach. donkeys have kicked away camera phones and sneezed in photographers’ faces.
human facial recognition has made a range of tasks easier, such as verifying passenger identity at train stations and catching criminals from surveillance footage. if animal recognition can reach similar levels of accuracy, it could help track herd health and reduce waste in agriculture, among the least digitized industries in the world, according to the mckinsey global institute.
farmers need to be able to tell animals apart to track their health and behavior. insurers hope facial recognition can help combat fraud by farmers. insurers have relied on farmers to identify animals by ear tags, but the animals often chew them off.
unfortunately, cow and pig faces aren’t so easy to tell apart with software originally designed for flatter human faces. and an array of four-legged creatures from china to ireland have been kicking up a fuss when technicians show up to take their pictures to better train the algorithms.
for better results, technicians recommend adding side profiles of animals, which have more distinguishing depth and features than those of humans. chen zenghui, innovationai’s product director, said he can discern how some donkeys have high cheekbones while others are blessed with sharp jawlines. he can tell cattle apart based on the slant of their eyes and the curve of their mouths.
the limited database of available animal faces, which companies say numbers in the hundreds of thousands—compared with millions for humans—slows the pace at which computers can learn. hence the messy and frustrating effort to collect more animal portraits.
“it’s not like you can tell a donkey to stand still and raise its chin up slightly,” said mr. chen. “that’s just impossible.”
facial-recognition software for animals, which is commercially available, doesn’t provide the same degree of speed and accuracy as similar technology for humans. the machines take longer to make a differentiation and can require high-definition photos or more profile shots from different angles.
animal faces require hundreds of reference points, many more than humans, to derive algorithms that can achieve similar levels of differentiation, scientists say.
“at the beginning, we weren’t even sure where exactly the focal point of a cow’s nose should be,” said huang xianjun, chief executive of beijing-based deepfinch ltd., a facial-recognition company. “the nostrils, or the tip, or maybe the area between the eyes?”
employees at cainthus ireland ltd., a dublin-based cow-recognition startup, learned to set up their video cameras out of reach of the cows’ tongues, which are as rough as sandpaper. the beasts have damaged cameras and bitten through the clips holding them in place.
“any way you can think that an animal with no hands could possibly cause problems with a camera—the cows did that,” said cainthus president david hunt. “and then they also established a whole bunch of other ways they could mess with the cameras that would never occur to any sane person ever.”
pigs, fleeter of foot and smoother of tongue, have resorted to other means of resistance.
“the pigs would move around so much that we had no choice but to have someone hold them down,” said huang hua, former vice president of marketing at yingzi technology ltd., an agricultural-technology firm in the southern chinese city of guangzhou. “sometimes it took dozens of attempts before we could get a good shot of one side of a pig’s face.”
animals are generally “not the camera-loving type,” said sun lu, product manager for deepfinch. during three days of trying to record videos of cows in 2017 and 2018, the beasts would head for the far corner whenever mr. sun and his team showed up.
mr. sun and his team got 12 different phone models and four special selfie sticks that can hold multiple phones. still, mr. sun had to chase a cow for five minutes before he could get a satisfactory shot.
“they are not good models, for sure,” he said.
on his first outing to photograph pigs, innovationai’s mr. deng made the error of arriving just before feeding time. he tried to hide as 30 hungry hogs turned on him. he slid and fell. “i had to throw out everything i was wearing,” he said.