salvacarrion commited on
Commit
decf4df
·
1 Parent(s): b7111bf

Drop redundant cbcl_aligned source (identical to cbcl by construction)

Browse files
Files changed (2) hide show
  1. README.md +11 -8
  2. data/train-00000-of-00001.parquet +2 -2
README.md CHANGED
@@ -22,7 +22,7 @@ classifiers).
22
 
23
  | Split | Rows | Faces / Non-faces |
24
  |---|---|---|
25
- | `train` | 109,406 | 104,858 / 4,548 |
26
  | `test` | 24,045 | 472 / 23,573 (CBCL benchmark) |
27
  | `negatives` | 29,879 | — / 29,879 (Caltech-256) |
28
 
@@ -47,18 +47,21 @@ negs = ds["negatives"]
47
 
48
  ## Face sources (train split)
49
 
50
- Each face source ships in **two paired variants at the same index** — same
51
- person, two crop styles:
52
 
53
- - `<src>` — loose portrait framing.
54
- - `<src>_aligned` — tight Viola-Jones crop (eyes pinned to fixed pixel
55
  positions, face fills the 48×48 frame).
56
 
 
 
57
  | `source` | Label | Count | Notes |
58
  |---|---|---|---|
59
- | `celeba` / `celeba_aligned` | 1 | 50,000 each | Aligned CelebA via manual landmarks; filtered for frontal pose |
60
- | `cbcl` / `cbcl_aligned` | 1 | 2,429 each | MIT CBCL #1, upsampled from native 19×19 (already canonically aligned) |
61
- | `cbcl` | 0 | 4,548 | CBCL non-faces matched-domain seed for stage 1 of cascade training |
 
62
 
63
  Picking `source == "cbcl"` from train returns **both faces and non-faces** —
64
  filter by `label` to pick.
 
22
 
23
  | Split | Rows | Faces / Non-faces |
24
  |---|---|---|
25
+ | `train` | 106,977 | 102,429 / 4,548 |
26
  | `test` | 24,045 | 472 / 23,573 (CBCL benchmark) |
27
  | `negatives` | 29,879 | — / 29,879 (Caltech-256) |
28
 
 
47
 
48
  ## Face sources (train split)
49
 
50
+ CelebA ships in **two paired variants at the same index** — same person,
51
+ two crop styles:
52
 
53
+ - `celeba` — loose portrait framing (hair and jaw visible).
54
+ - `celeba_aligned` — tight Viola-Jones crop (eyes pinned to fixed pixel
55
  positions, face fills the 48×48 frame).
56
 
57
+ CBCL is already canonically aligned, so it has a single variant.
58
+
59
  | `source` | Label | Count | Notes |
60
  |---|---|---|---|
61
+ | `celeba` | 1 | 50,000 | Loose portrait. Filtered for frontal pose using the manual CelebA landmarks. |
62
+ | `celeba_aligned` | 1 | 50,000 | Same 50,000 faces tightly warped to CBCL-matched geometry. |
63
+ | `cbcl` | 1 | 2,429 | MIT CBCL #1, upsampled from native 19×19. Already canonically aligned. |
64
+ | `cbcl` | 0 | 4,548 | CBCL non-faces — matched-domain seed for stage 1 of cascade training. |
65
 
66
  Picking `source == "cbcl"` from train returns **both faces and non-faces** —
67
  filter by `label` to pick.
data/train-00000-of-00001.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:7549849f0c4a24ade8b9297cd9230103e48e60badb4d1e279e44ef19b57a73cb
3
- size 190040507
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:13aeb0a839ee7bb9e137ae4cfb543c52966044893574daf30dfae71064dd50b6
3
+ size 186679425