Skip to content

Latest commit

 

History

History
88 lines (73 loc) · 5.58 KB

02_data_analysis.md

File metadata and controls

88 lines (73 loc) · 5.58 KB

Data Analysis

TODO: Add image resolution analysis for both datasets.

Note: FashionFail is not intended to replace the Fashionpedia dataset directly. Rather, they target different domains—Fashionpedia focuses on 'in-the-wild' images showcasing multiple apparel items on a human body, whereas FashionFail concentrates on singular items commonly found in the e-commerce domain. However, FashionFail can still be utilized to assess the robustness of trained models across the e-commerce domain, making it valuable for evaluating model performance in real-world e-commerce settings.

Table: The class distribution in Fashionpedia-train and FashionFail-train.

class Fashionpedia# Fashionpedia% FashionFail# FashionFail#
shoe 46374 30.31% 441 32.81%
dress 18739 12.25% 29 2.16%
top, t-shirt, sweats. 16548 10.82% 71 5.28%
pants 12414 8.11% 53 3.94%
jacket 7833 5.12% 45 3.35%
bag, wallet 7217 4.72% 146 10.86%
shirt, blouse 6161 4.03% 57 4.24%
skirt 5046 3.30% 26 1.93%
glasses 4855 3.17% 62 4.61%
tights, stockings 4326 2.83% 39 2.90%
headband, head c. 3470 2.27% 45 3.35%
watch 3389 2.22% 99 7.37%
coat 3124 2.04% 12 0.89%
shorts 2756 1.80% 52 3.87%
sock 2582 1.69% 22 1.64%
hat 2518 1.65% 88 6.55%
glove 1385 0.91% 12 0.89%
scarf 1374 0.90% 2 0.15%
cardigan 1107 0.72% 1 0.07%
jumpsuit 922 0.60% 16 1.19%
vest 719 0.47% 23 1.71%
umbrella 135 0.09% 3 0.22%
TOTAL 152994 100% 1344 100%

Comparison with Fashionpedia

Fashionpedia(FP) has the following splits (with number of images);
Train (45,623), Validation (1,158), and Test (2,044).
FashionFail(FF) has the following splits (with number of images);
Train (1,344), Validation (150), and Test (1,001).

The following table shows the number of samples in each split for each class. For example, FF-train contains 57 samples of shirt, blouse class. As FF has exactly 1 annotation per image, this means there are 57 images (out of 1344) of shirt, blouse class inside FF-train. On the contrary, FP has many annotations per image — FP-val has a total of 4481 annotations for 1158 images.

Class / Split FF-train FF-val FF-test FP-train⬇ FP-val
23 shoe 441 49 328 46374 1566
10 dress 29 3 22 18739 508
1 top,t-shirt,sweats 71 8 53 16548 477
6 pants 53 6 40 12414 314
4 jacket 45 5 34 7833 183
24 bag, wallet 146 16 109 7217 214
0 shirt, blouse 57 6 42 6161 102
8 skirt 26 3 20 5046 162
13 glasses 62 7 46 4855 130
21 tights, stockings 39 4 29 4326 122
15 headband,head c. 45 5 33 3470 109
18 watch 99 11 73 3389 84
9 coat 12 1 8 3124 104
7 shorts 52 6 39 2756 106
22 sock 22 3 17 2582 87
14 hat 88 10 65 2518 74
17 glove 12 2 10 1385 31
25 scarf 2 0 1 1374 48
3 cardigan 1 0 1 1107 12
11 jumpsuit 16 2 12 922 21
5 vest 23 3 18 719 22
26 umbrella 3 0 1 135 5
TOTAL 1344 150 1001 152994 4481