Open-vocabulary object detection (OVOD) aims to detect known and unknown objects in the open world by leveraging text prompts. Benefiting from the emergence of large-scale vision--language pre-trained models, OVOD has demonstrated strong zero-shot generalization capabilities. However, when dealing with camouflaged objects, the detector often fails to distinguish and localize objects because the visual features of the objects and the background are highly similar. To bridge this gap, we construct a benchmark named OVCOD-D by augmenting carefully selected camouflaged object images with fine-grained textual descriptions. Due to the limited scale of available camouflaged object datasets, we adopt detectors pre-trained on large-scale object detection datasets as our baseline methods, as they possess stronger zero-shot generalization ability. In the specificity-aware sub-descriptions generated by multimodal large models, there still exist confusing and overly decorative modifiers. To mitigate such interference, we design a sub-description principal component contrastive fusion strategy that reduces noisy textual components. Furthermore, to address the challenge that the visual features of camouflaged objects are highly similar to those of their surrounding environment, we propose a specificity-guided regional weak alignment and dynamic focusing method, which aims to strengthen the detector's ability to discriminate camouflaged objects from background. Under the open-set evaluation setting, the proposed method achieves an AP of 56.4 on the OVCOD-D benchmark.
Figures & tables
Figure 1: We perform zero-shot detection with the YOLO-World-M [ 2 ] on both the LVIS [ 10 ] dataset and our OVCOD-D dataset. By comparing the AP of the overlapping categories across the two datasets, we observe a substantial performance decline on OVCOD-D, indicating that open-vocabulary detectors face significant challenges when dealing with camouflaged objects.
Figure 2: We conducted a statistical analysis of per-class instance counts in the OVCOD-D dataset. In the bar chart, blue bars denote base classes and red bars denote novel classes, revealing a pronounced long-tailed distribution of class frequencies. Additionally, using the fine-grained textual descriptions associated with each category, we selected 25 categories characterized by higher lexical richness to construct a multidimensional quality-analysis heatmap. The horizontal axis reports, in sequence, lexical diversity, average tokens, unique words, average unique-word ratio, and the standard deviation of sentence length.
Figure 3: Overall architecture of the proposed specificity-driven open-vocabulary camouflaged object detector. Fine-grained textual sub-descriptions are encoded by a text encoder and decorrelated via SVD, then refined through an adapter and integrated with visual object embeddings in the contrastive fusion module. The image is processed by a lightweight YOLO‑style backbone and a PAN to extract multi‑scale features, which are transformed into object embeddings and fed in parallel into the SF‑GLU and the box head.
Figure 4: Construction pipeline of OVCOD-D dataset. We extend COD10K-D, NC4K-D, and clean CAMO-D with YOLO-style detection labels and an additional red imported fire ant nest subset, then reorganize them into 40 base and 47 novel classes. Qwen3-VL-Plus generates fine-grained image descriptions from which we derive a semantic prompt library for open-vocabulary camouflaged object detection.
Method
Backbone
Params
Pre-train
AP
AP 50
AP 75
AP m
AP l
GLIP-T [ 20 ]
Swin-T
232M
O365,GoldG
22.6
31.9
28.2
-
-
GLIPv2-T [ 44 ]
Swin-T
232M
O365,GoldG
25.6
36.4
29.7
-
-
Grounding DINO-T [ 25 ]
Swin-T
172M
O365,GoldG
34.8
43.9
37.7
-
-
YOLO-World-S [ 2 ]
YOLOv8-S
77M
O365,GoldG
38.5
58.0
41.4
18.7
41.1
YOLO-World-M [ 2 ]
YOLOv8-M
92M
O365,GoldG
43.3
62.3
46.5
24.9
46.1
YOLO-World-L [ 2 ]
YOLOv8-L
110M
O365,GoldG
45.7
63.2
48.9
22.9
48.4
Table 1: Comparison with classical open-vocabulary object detectors on the OVCOD-D dataset is conducted under an open-set setting, where all models are fine-tuned on the base classes and evaluated on the union of base and novel classes.
Method
Pre-train
AP
AP 50
AP 75
FPS
Training from scratch.
YOLOv8-S [ 34 ]
×
18.6
26.5
19.0
-
YOLOv8-M [ 34 ]
×
21.7
31.8
22.6
-
YOLOv8-L [ 34 ]
×
23.7
32.0
23.4
-
Fine-tuned w/ DuVL-PAN and sub-descriptions
YOLO-World-S [ 2 ]
O+G
39.3
57.3
41.3
35
Table 2: Comparisons under the closed-set setting are conducted to evaluate the performance of classical object detectors and open-vocabulary object detectors on the OVCOD-D dataset. The FPS is measured on a single NVIDIA RTX 4090 using the raw trained checkpoints without TensorRT, re-parameterization, or other lightweight optimizations.
Fusion Method
AP
AP 50
AP 75
Orthogonal+Contrastive
46.5
70.1
48.9
Contrastive
47.5
70.6
49.6
Orthogonal
44.0
66.4
45.6
Table 3: Ablation study on sub-description fusion strategies for open-vocabulary camouflaged object detection on OVCOD-D dataset. The coverage auxiliary loss weight is uniformly set to 5, and SF-GLU is disabled for all variants to ensure fair comparison.
Dynamic Focusing Module
AP
AP 50
AP 75
w/o GLU
47.5
70.6
49.6
GLU
46.7
69.7
49.8
SwiGLU
47.1
70.3
49.2
SF-GLU
48.1
70.7
50.3
Table 4: Ablation study on SF-GLU module for open-vocabulary camouflaged object detection on the OVCOD-D dataset. All variants adopt the contrastive fusion framework with coverage auxiliary loss set to 5.
Weight for auxiliary loss
AP
AP 50
AP 75
0
47.3
71.0
49.5
2
47.2
70.5
48.7
5
47.5
70.6
49.6
7
46.3
69.3
49.0
10
46.6
69.7
49.0
Table 5: Ablation study on different weights of coverage-based auxiliary loss for open-vocabulary camouflaged object detection on the OVCOD-D dataset. All variants adopt the same contrastive fusion framework without using the SF-GLU module.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Metric
Accuracy
Relevance
Hallucination-free
Qualified category
Rate (%)
98.15
99.77
92.61
100.00
Appendix
Table A1 : Human evaluation of generated textual descriptions.
Method
AP
AP50
AP75
SINet-V2 [ 5 ]
40.2
69.3
39.4
FSPNet [ 16 ]
47.9
76.2
49.4
CamoFormer [ 42 ]
55.6
80.2
59.0
HDPNet [ 11 ]
56.3
81.5
59.6
SDDF-L (Ours)
56.4
76.4
60.7
Appendix
Table A2 : Comparison with SOTA COD methods on OVCOD-D.
Method
APmild
APmod
APsevere
SDDF-S
64.2
49.8
39.0
Appendix
Table A3 : Multi-level camouflability performance of SDDF-S on OVCOD-D.
SVD Configuration
AP
AP 50
AP 75
w/o SVD
44.8
67.4
46.8
w/ SVD (5)
46.5
69.8
48.3
w/ SVD (10)
46.8
69.5
49.2
w/ SVD (15)
46.1
69.2
48.4
Appendix
Table A4 : Ablation study on different maximum number of components obtainable from SVD on the OVCOD-D dataset. All model variants adopt the fusion approach defined in Equation A1 . The coverage auxiliary loss and SF-GLU are disabled.
Conv Layers
AP
AP50
AP75
1
48.1
70.7
50.3
2
47.1
69.6
49.8
4
46.0
68.7
48.6
Appendix
Table A5 : Ablation study on the number of convolutional layers.
Figure A1 : Quantitative comparison is conducted via visualization of the detection bounding boxes.
Figure A2 : Quantitative comparisons are further conducted by visualizing heatmap representations of the response differences in feature maps with and without the proposed SF-GLU. Specifically, these feature maps are extracted from the PAN outputs.