Skelibras: a large annotated skeleton dataset of dynamic Libras signs comprised of unsegmented and segmented conversations.
The directory structure is simple with each folder following the name of the sign. And each sign follows this nomenclature below:
- VideoID---SignName---BeginInMiliseconds---EndInMiliSeconds---View---SignalyID.csv
The signs names follow the guia corpus guide:
All CSVs have the keys corresponding to the extracted joints, hand and body. All indices correspond to frames. And each joint is a string from a numpy.array, which can be mapped back to numpy.array as follows:
def parse_npy_vec_str(str_array_like):
if not isinstance(str_array_like, str):
return str_array_like
res = str_array_like[1:len(str_array_like) - 1].split(' ')
recovered_np_array = []
for r in res:
if r == '' or ' ' in r:
continue
f = float(r)
recovered_np_array.append(f)
return np.array(recovered_np_array)
sign = sign.applymap(parse_npy_vec_str)