Industrial shampoo-bottle surface defect detection must operate across many stock keeping units (SKUs), where differences in bottle geometry, material, label layout, reflectance, and imaging conditions can substantially alter the appearance of the same defect. Existing industrial defect datasets rarely provide both instance-level defect annotations and product-variant identities, making it difficult to study SKU-aware detection under realistic manufacturing conditions. We introduce Bottle-SKU, a large-scale multi-SKU benchmark for shampoo bottle surface defect detection. Bottle-SKU contains 13,119 images, 20,419 instance-level bounding boxes, seven defect categories, and 48 bottle SKUs, enabling evaluation of both shared defect recognition and SKU-specific appearance variation. To address the challenge of SKU-specific appearance variation, we propose Prototype-Guided Dynamic Domain Modulation (ProDMM). ProDMM constructs a visual prototype for each SKU from training images and predicts a soft SKU distribution for each input. The resulting prototype-conditioned representation dynamically generates low-rank Query/Value modulation for the visual backbone and residual conditioning for defect-text-to-visual attention, while preserving the pretrained Grounding DINO detection pathway. The model is trained jointly with detection, SKU-domain classification, and modulation-regularization objectives. Experiments on Bottle-SKU and the public GC10-DET-SKU metallic surface defect dataset demonstrate the effectiveness of prototype-guided domain conditioning for industrial surface defect detection.
Figures & tables
Figure 1: Overview of the proposed cross-SKU defect detection benchmark and method. ( Left ) Appearance gap between training and deployment. Shampoo bottles from different SKUs vary substantially in appearance and material, while defects such as scratches and wrinkles remain visually similar. ( Middle ) Task setting . We adopt an SKU-level deployment protocol, training on a set of seen SKUs ( SKU-1, SKU-2, … ) and testing on unseen SKUs ( SKU-a, SKU-b, … ), which emulates the real-world scenario where a production line switches to a new product. ( Right ) (a) Conventional visual-language methods feed visual and text features directly into the detection head. (b) Our prototype-guided modulation injects domain shift into the text and visual branches, enabling robust detection on unseen SKUs.
Figure 2: The overall architecture of ProDMM. Offline SKU prototypes condition dynamic QLer/VLer residuals in the Swin backbone and residuals for defect Domain Cross Attention (DCA). DCA updates only the text memory before the original GroundingDINO Feature Enhancer. The bottom-right inset illustrates a DMT block.
Dataset
Application
Number
Category
Appearance Complexity
Inspection Level
Image Resolution Pixels
NEU-DET Song and Yan (2013)
Hot-rolled steel strip
1,800
6
Low
Local
120×120
GC10-DET Lv et al. (2020)
Steel strip
2,294
10
Medium
Local
2048×1000
PVMulti-Defect Li et al. (2023b)
Solar photovoltaic panels
307
5
Medium
Local
5800×3504
HRIPCB Huang et al. (2020a)
HRIPCB
1,386
6
High
Full object
600×600
Severstal Severstal (2019)
Steel strip
18,074
4
Low
Local
1600×256
DAGM Wieler et al. (2007)
Textured background
16,100
6
High
Local
800×800
Table 1: Comparison of public datasets for surface defect detection. Appearance complexity is qualitatively assessed according to surface texture, background pattern, object geometry, material reflectance, and intra-class appearance variation.
Index
Algorithm
Publish
Bottle-SKU
GC10-DET-SKU
Params
AP
AP50
AP75
P
R
F1
AP
AP50
AP75
P
R
F1
Lightweight Detector
1
Faster R-CNN Ren et al. (2017)
TPAMI 2017
9.6
28.2
3.6
32.6
60.4
42.3
20.8
43.6
19.0
29.7
60.4
39.8
41.38M
2
Cascade R-CNN Cai and Vasconcelos (2018)
CVPR 2018
10.6
29.8
4.7
42.2
55.9
48.1
26.0
51.5
23.1
44.5
62.6
52.0
69.17M
3
FCOS Tian et al. (2019)
ICCV 2019
9.5
28.2
4.1
56.7
40.9
47.5
18.3
41.7
15.1
59.3
35.3
44.2
32.13M
4
GFL Li et al. (2020)
NeurIPS 2020
11.1
31.3
5.0
30.4
63.2
41.1
24.8
50.6
23.0
31.4
61.2
41.5
32.27M
Table 2: Comparison on Bottle-SKU and the GC10-DET-SKU pseudo-domain protocol. All values are percentages.
No.
Prototype
DMT
DCA
AP
AP 50
AP 75
AP S
AP M
AP L
01
✗
✗
✗
15.1
39.8
7.9
2.5
9.1
16.6
02
✗
✓
✗
15.5
40.7
8.1
2.2
8.9
17.1
03
✓
✓
✗
15.8
41.6
8.0
2.2
8.4
17.4
04
✓
✓
✓
16.0
40.8
8.9
2.3
8.7
17.9
Table 3: Component Analysis.
Table 6
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Examples of bottle images from different SKUs in our dataset.
Figure 4: Examples of defects appearing on bottle surfaces.
Split
#SKU
#Images
#Instances
Ins./Img.
Train
38
10,495
16,194
1.54
Test
10
2,624
4,225
1.61
Total
48
13,119
20,419
1.56
Appendix
Table 6: Statistics of the Bottle-SKU dataset.
Figure 5: Distribution of annotated defect instances across Bottle-SKU product variants. The rows are divided into training SKUs and held-out test SKUs. Color intensity represents log10(1+#instances) , and the bars on the left indicate the number of images for each SKU.
Figure 6: A t-SNE visualization of the features extracted from the training set for the 38 SKU classes.
No.
Methods
AP
AP 50
AP 75
AP S
AP M
AP L
01
LoRA rank=1
15.5
39.8
8.1
2.8
9.8
17.2
02
LoRA rank=64
15.7
40.4
8.1
3.3
9.5
17.3
03
LoRA rank=128
15.8
39.7
8.8
3.2
9.2
17.3
04
AdaptFormer
15.7
40.4
8.3
3.8
9.1
17.5
05
Ours
16.0
40.8
8.9
2.3
8.7
17.9
Appendix
Table 7: Comparison with static parameter-efficient fine-tuning baselines.
No.
Methods
AP
AP 50
AP 75
AP S
AP M
AP L
01
0.25
15.4
41.0
7.6
2.0
8.8
17.1
02
0.5
15.6
39.9
8.4
3.0
10.0
17.2
03
1.0
16.0
40.8
8.9
2.3
8.7
17.9
Appendix
Table 8: Parameter Analysis of the Domain Classifier Loss Weight.
Figure 7: Comparison of Ground Truth, ProDMM, DOSOD, PET-DINO, WeDetect and YOLO-World.
Industrial inspection uses sensing, computer vision, and robotics to detect, measure, and explain defects and verify product or infrastructure quality.