Fabric Defect Detection — Tianchi Guangdong Round 1
Detects 20 classes of fabric defects in 2446 × 1000 industrial loom images.
mAP@0.5 = 0.4992 on a held-out fold, versus 0.442 best published on the same 20-class benchmark (Jin et al. 2025, Appl. Sci. 15:3228, Table 4).
The core idea
Every image is 2446 × 1000 px and 25% of defects are smaller than 32² px — the median small defect is 20 × 27 px.
Standard practice letterboxes to 640 × 640. That scales by 640/2446 = 0.262, turning the median small defect into 5.3 × 7.2 px — smaller than a single stride-8 grid cell. It cannot be assigned a positive training sample, so it receives no gradient at all. This is why every published model collapses on this dataset (best 0.442) while the same architectures reach 0.71–0.82 on TILDA, whose 416 × 416 images lose nothing to downscaling.
Solution: three full-height 1000 px strips at native resolution (the defect stays 20 × 27 px) plus a global view for defects wider than any strip, fused with Weighted Box Fusion.
Results
| mAP@0.5 | AP small | AP large | |
|---|---|---|---|
| 640² control | 0.3905 | 0.2129 | 0.3483 |
| tiled | 0.4464 | 0.4374 | 0.3774 |
| + seam-aware fusion | 0.4821 | 0.4393 | 0.4273 |
| + hflip TTA | 0.4992 | 0.4620 | 0.4092 |
Small-object AP rose +117%. The final +0.053 required no retraining — identical weights, changed inference only.
Seam-aware fusion
A 1000 px strip cannot represent a 2446 px box, so it emits a truncated stump. Weighted Box Fusion then averages three stumps against the global branch's one correct box and drags the result inward. Suppressing strip predictions that touch an interior seam (as opposed to the real image border) recovers +0.036 mAP for free. Toggle it in the sidebar.
Notes
- Input must be exactly 2446 × 1000 — other sizes are rejected rather than silently mis-mapped.
- The free CPU tier runs 4 views per image: expect 10–25 s. The same configuration takes 0.11 s on an RTX 2000 Ada.
Credits
Graduation project supervised by Dr. Tarek Ghonim.
Khaled Awad · Hazem Shokr · Hazem Mohamed · Bassem Salah · Hossam Elnagar
Dataset: Alibaba Tianchi Guangdong Industrial Intelligent Manufacturing — Cloth Flaw Detection, Round 1.