Found:
from PyPDF2 import PdfFileReader
reader = PdfFileReader(open("1707.09725.pdf", 'rb'))
print(reader.outlines)
gives:
[{'/Title': '1 Introduction', '/Left': 99.213, '/Type': '/XYZ', '/Top': 742.911, '/Zoom': ..., '/Page': IndirectObject(513, 0)},
{'/Title': '2 Convolutional Neural Networks', '/Left': 99.213, '/Type': '/XYZ', '/Top': 742.911, '/Zoom': ..., '/Page': IndirectObject(554, 0)}, [{'/Title': '2.1 Linear Image Filters', '/Left': 99.213, '/Type': '/XYZ', '/Top': 486.791, '/Zoom': ..., '/Page': IndirectObject(554, 0)},
{'/Title': '2.2 CNN Layer Types', '/Left': 70.866, '/Type': '/XYZ', '/Top': 316.852, '/Zoom': ..., '/Page': IndirectObject(580, 0)},
[{'/Title': '2.2.1 Convolutional Layers', '/Left': 99.213, '/Type': '/XYZ', '/Top': 562.722, '/Zoom': ..., '/Page': IndirectObject(608, 0)},
{'/Title': '2.2.2 Pooling Layers', '/Left': 99.213, '/Type': '/XYZ', '/Top': 299.817, '/Zoom': ..., '/Page': IndirectObject(654, 0)},
{'/Title': '2.2.3 Dropout', '/Left': 99.213, '/Type': '/XYZ', '/Top': 742.911, '/Zoom': ..., '/Page': IndirectObject(689, 0)},
{'/Title': '2.2.4 Normalization Layers', '/Left': 99.213, '/Type': '/XYZ', '/Top': 193.779, '/Zoom': <PyPDF2.generic.NullObject object at 0x7fbe49d14350>, '/Page': IndirectObject(689, 0)}]