Summary:
`load_obj` recognizes vertex, texture, normal and face lines with `line.startswith("v ")`, `"vt "`, `"vn "` and `"f "`. The OBJ format separates fields with any whitespace, and some exporters write a tab after the keyword. Those lines are skipped without any error, so the mesh silently loses vertices and faces:
```python
from io import StringIO
from pytorch3d.io import load_obj
verts, faces, _ = load_obj(StringIO("v\t0 0 0\nv 1 0 0\nv 0 1 0\nf\t1 2 3\n"))
verts.shape, faces.verts_idx.shape
# before: (2, 3), (0, 3) after: (3, 3), (1, 3)
```
This dispatches on the first token of the already split line (`tokens[:1] == ["v"]`), the same way the `usemtl` branch already does.
Pull Request resolved: https://github.com/facebookresearch/pytorch3d/pull/2061
Test Plan: added `test_load_obj_tab_separated` to `tests/test_io_obj.py`, using tabs in `v`, `vt`, `vn` and `f` lines. It fails on main and passes with the change. I ran `tests/test_io_obj.py` on Windows without the C++ extension: the same 19 tests fail before and after (they reopen a `NamedTemporaryFile` or need `_C`), and the rest pass. flake8 is clean on the changed lines.
Reviewed By: MichaelRamamonjisoa
Differential Revision: D124656223
Pulled By: bottler
fbshipit-source-id: ab0ef3dc0f3c998c44874e6a416e6a9649f29378
8b3ba6ede4