Reviving the Discarded High-Resolution Feature for Transformer-Based Tiny Object Detection

Zehao Wang; Si Wu; Guiping Cao; Xin Li; Yaowei Wang

cs.CVNov 5, 2026

Reviving the Discarded High-Resolution Feature for Transformer-Based Tiny Object Detection

Authors: Zehao Wang, Si Wu, Guiping Cao, Xin Li, Yaowei Wang

Organizations: South China University of Technology, Guangzhou, China · Pengcheng Laboratory, Shenzhen, China · Harbin Institute of Technology, Shenzhen, China

Abstract

Tiny object detection remains challenging because tiny instances occupy only a few pixels, and their fine-grained details are easily lost during feature downsampling. Although the highest-resolution feature preserves rich local details for tiny objects, DETR-like detectors often discard it due to its large computational cost, while directly reintroducing this feature also brings dense background noise. To address this problem, we propose ReDHF, a novel framework that Revives the Discarded High-resolution Feature for tiny object detection. ReDHF has three cooperative components. Cross-Scale Deformable Fusion (CSDF) first constructs a semantically enhanced highest-resolution feature and injects its fine-grained details into the shallow feature under a moderate cost. Ellipse-Guided Shallow Supervision (EGSS) then guides the enhanced feature to learn more reliable classification and localization priors by applying geometry-adaptive auxiliary supervision. Full-Scale Query Interaction (FSQI) further enables decoder queries to access both low-level details and high-level semantic information, allowing them to aggregate richer cues for tiny objects. Extensive experiments show that ReDHF consistently improves multiple baselines across different DETR variants and achieves state-of-the-art performance on the evaluated benchmarks. The best ReDHF variants surpass their baselines by +1.4 AP on AI-TOD-v2, +1.6 AP on VisDrone, and +1.5 AP on SODA-D, with especially notable gains on tiny and small objects.

Figures & tables

Figure 1: (a) Tiny objects may occupy only a few pixels, making their visual evidence extremely weak. (b) Very tiny objects can become nearly invisible on stride-4 feature map in the feature downsampling pipeline. (c) The ratio of ultra-tiny objects (short side $\leq 4$ pixels) is much higher in tiny object datasets such as AI-TOD-v2, VisDrone, and SODA-D than in COCO.

Figure 2: Overview of ReDHF. CSDF revives the discarded highest-resolution feature $C_{1}$ as a semantically enhanced feature $P_{1}$ and injects its fine-grained details into $P_{2}$ to obtain the detail-enhanced shallow feature $P_{2}^{\prime}$ . EGSS further supervises $P_{2}^{\prime}$ with ellipse-guided auxiliary supervision during training, while FSQI exposes $P_{1}$ , $P_{2}^{\prime}$ , and deeper features to decoder queries for final detection.

CNN-based models
Method	#Params.	GFLOPs	AP	AP 50	AP 75	AP vt	AP t	AP s	AP m
Cascade R-CNN [ 1 ]	68.9M	88.7	15.1	34.2	11.2	0.1	11.5	26.7	38.5
DetectoRS [ 32 ]	134.8M	$\approx$ 160	16.1	35.5	12.5	0.1	12.6	28.3	40.0
DetectoRS w/ RFLA [ 47 ]	–	–	24.8	55.2	18.5	9.3	24.8	30.3	38.2
YOLOv12-X [ 42 ]	59.3M	185.4	16.1	33.5	13.4	5.4	17.0	21.4	28.5
Transformer-based models

Table 1: Experiments on AI-TOD-v2 test split. All models are evaluated with 800×800 input resolution. We compare ReDHF with both CNN-based and Transformer-based detectors.

\dagger

denotes our re-implementation of the corresponding baseline.

\star

denotes the average GFLOPs measured over the whole split when the computational cost varies across images.

Baseline	VisDrone AP	VisDrone AP s	SODA-D AP	SODA-D AP eS
DINO-4scale [ 52 ]	34.3 $\rightarrow$ 35.4 (+1.1)	25.6 $\rightarrow$ 27.4 (+1.8)	28.6 $\rightarrow$ 30.1 (+1.5)	10.5 $\rightarrow$ 12.6 (+2.1)
DQ-DETR [ 15 ]	35.7 $\rightarrow$ 37.1 (+1.4)	28.2 $\rightarrow$ 29.6 (+1.4)	34.4 $\rightarrow$ 35.5 (+1.1)	17.0 $\rightarrow$ 18.5 (+1.5)
Dome-DETR-S [ 14 ]	35.9 $\rightarrow$ 37.5 (+1.6)	28.3 $\rightarrow$ 30.3 (+2.0)	35.4 $\rightarrow$ 36.8 (+1.4)	17.6 $\rightarrow$ 19.6 (+2.0)
Dome-DETR-M [ 14 ]	38.9 $\rightarrow$ 40.2 (+1.3)	31.4 $\rightarrow$ 33.1 (+1.7)	37.2 $\rightarrow$ 38.4 (+1.2)	18.6 $\rightarrow$ 20.4 (+1.8)
Dome-DETR-L [ 14 ]	39.0 $\rightarrow$ 40.6 (+1.6)	31.5 $\rightarrow$ 33.1 (+1.6)	37.6 $\rightarrow$ 38.7 (+1.1)	19.7 $\rightarrow$ 21.4 (+1.7)
Avg. gain	+1.4	+1.7	+1.3	+1.8

Table 2: Experiments on VisDrone validation split and SODA-D test split. Each entry reports baseline

\rightarrow

ReDHF, with the gain over the corresponding baseline shown in parentheses. AP s is used as the tiny-object metric for VisDrone, and AP eS is used for SODA-D.

Table 5

Figure 3: Qualitative comparison on the AI-TOD-v2 test split. Each pair compares a baseline detector with its ReDHF variant on the same scene. The green , red , yellow , and blue boxes represent TP, FP, FN with ground-truth box short side $\leq 4$ pixels, and the remaining FN, respectively. Compared with the baselines, ReDHF produces cleaner predictions and helps recover more tiny objects, with more green boxes and fewer red/yellow/blue boxes. Class labels are omitted for clarity.

Figure 4: Feature visualization and response analysis of CSDF. We compare $C_{1}$ vs. $P_{1}$ and $P_{2}$ vs. $P_{2}^{\prime}$ . The bottom-left statistics report $\mu_{\text{in}}$ , $\mu_{\text{out}}$ , and FBR defined in Eq. 7 . After CSDF enhancement, $P_{1}$ and $P_{2}^{\prime}$ show stronger response concentration on GT regions.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

CNN-based models
Method	#Params.	GFLOPs	Image Scale	AP	AP 50	AP 75	AP s	AP m	AP l
YOLOv11 [ 17 ]	25.3M	86.6	640×640	27.9	–	–	18.1	40.9	53.1
YOLOv12 [ 42 ]	26.3M	88.6	640×640	29.2	–	–	18.7	41.8	51.7
RFLA [ 47 ]	69.2M	162.0	800×800	24.3	–	–	14.4	35.9	45.8
YOLC [ 22 ]	–	–	1024×640	39.6	–	–	32.9	49.3	56.2
CFINet [ 51 ]	41.4M	291.9	1200×1200	23.9	–	–	16.3	34.7	57.4

Appendix

Table 6: Experiments on VisDrone val split. We compare ReDHF with both CNN-based and Transformer-based detectors.

\dagger

denotes our re-implementation of the corresponding baseline.

\star

denotes the average GFLOPs measured over the whole split when the computational cost varies across images.

CNN-based models
Method	#Params.	GFLOPs	Image Scale	AP	AP 50	AP 75	AP eS	AP rS	AP gs	AP N
Faster RCNN [ 35 ]	41.2M	292.3	1200×1200	28.9	59.7	24.2	13.9	25.6	34.3	43.2
Cascade RCNN [ 1 ]	68.9M	320.1	1200×1200	31.2	59.9	27.8	14.1	27.5	37.1	46.9
RFLA [ 47 ]	41.2M	292.1	1200×1200	29.7	60.2	25.2	13.2	26.9	35.4	44.6
Transformer-based models
Deformable-DETR [ 59 ]	35.2M	739.1	1200×1200	19.2	44.8	13.7	6.3	15.4	24.9	34.2

Appendix

Table 7: Experiments on SODA-D test split. We compare ReDHF with both CNN-based and Transformer-based detectors.

\dagger

denotes our re-implementation of the corresponding baseline.

\star

denotes the average GFLOPs measured over the whole split when the computational cost varies across images.

Model	Backbone	Encoder	Decoder	Others	Whole
DINO-4scale [ 52 ]	104.5	203.9	34.9	11.1	354.4
DINO-4scale w/ direct $C_{1}$	104.5 (–)	2709.5 (+2505.6)	161.0 (+126.1)	82.7 (+71.6)	3057.7 (+2703.3)
DINO-4scale w/ ReDHF	104.5 (–)	251.2 (+47.3)	40.7 (+5.8)	11.4 (+0.3)	407.8 (+53.4)

Appendix

Table 8: Computational cost comparison on DINO-4scale. We report GFLOPs of the backbone, encoder, decoder, other parts, and the whole model. “Direct

C_{1}

” denotes directly adding the raw highest-resolution feature into the encoder and decoder. “Others” includes components not counted in the backbone, encoder, or decoder.

Model	Input	GFLOPs	AP	AP vt
DINO-4scale	800	354.4	25.1	12.1
DINO-4scale	864	412.4	25.4	12.4
DINO-4scale w/ ReDHF	800	407.8	26.0	13.8
Dome-DETR-L	800	358.7	34.6	19.6
Dome-DETR-L	880	481.1	35.0	20.0
Dome-DETR-L w/ ReDHF	800	415.3	35.7	20.8

Appendix

Table 9: Compute-matched comparison with input-resolution scaling on AI-TOD-v2. We increase only the input resolution of the baselines while keeping the detector and training settings unchanged. The best result within each baseline family is highlighted in bold.

Dataset	Variant	Tiny AP $\uparrow$	Latency (ms) $\downarrow$	Avg. memory (GiB) $\downarrow$	Peak memory (GiB) $\downarrow$
AI-TOD-v2	Dome-DETR-S	17.5 $\rightarrow$ 18.9	409.6 $\rightarrow$ 507.6	0.30 $\rightarrow$ 0.56	0.63 $\rightarrow$ 1.72
	Dome-DETR-M	18.9 $\rightarrow$ 20.7	450.0 $\rightarrow$ 578.0	0.39 $\rightarrow$ 0.65	0.77 $\rightarrow$ 1.86
	Dome-DETR-L	19.6 $\rightarrow$ 20.8	487.8 $\rightarrow$ 655.2	0.47 $\rightarrow$ 0.76	0.86 $\rightarrow$ 1.99
VisDrone	Dome-DETR-S	28.3 $\rightarrow$ 30.3	454.5 $\rightarrow$ 584.1	0.27 $\rightarrow$ 0.46	0.63 $\rightarrow$ 1.72
	Dome-DETR-M	31.4 $\rightarrow$ 33.1	512.8 $\rightarrow$ 680.1	0.36 $\rightarrow$ 0.57	0.77 $\rightarrow$ 1.86
	Dome-DETR-L	31.5 $\rightarrow$ 33.1	543.5 $\rightarrow$ 751.8	0.47 $\rightarrow$ 0.69	0.88 $\rightarrow$ 2.01

Appendix

Table 10: Practical inference cost of ReDHF on AI-TOD-v2 and VisDrone. Wall-clock latency and GPU memory are measured on a single Tesla V100 with batch size 1 under identical baseline and ReDHF settings. Each entry reports baseline

\rightarrow

ReDHF. Tiny AP denotes AP vt on AI-TOD-v2 and AP s on VisDrone.

Dataset	Smallest objects	Intermediate objects	Larger objects
AI-TOD-v2	AP vt +1.48	AP t +0.86	AP s +0.40, AP m +0.00
VisDrone	AP s +1.70	AP m +0.86	AP l +0.32
SODA-D	AP eS +1.82	AP rS +1.26	AP gS +0.82, AP N +0.40

Appendix

Table 11: Scale-wise performance gains of ReDHF averaged over the five evaluated DETR variants. The dataset-specific metrics are ordered from smaller to larger object scales.

Model	AP	AP s	AP m	AP l
DINO-4scale	49.0	32.0	52.3	63.0
DINO-4scale w/ ReDHF	49.6 (+0.6)	32.6 (+0.6)	53.1 (+0.8)	63.5 (+0.5)

Appendix

Table 12: Generalization of ReDHF on COCO using DINO-4scale. Both models use the same 12-epoch training recipe.

Variant	Sparse AP s	Sparse AP m	Sparse AP l	Dense AP s	Dense AP m	Dense AP l
Dome-DETR-S	28.24 $\rightarrow$ 30.31 ( +2.07 )	47.09 $\rightarrow$ 48.91 ( +1.82 )	60.60 $\rightarrow$ 63.12 ( +2.52 )	28.11 $\rightarrow$ 30.58 ( +2.47 )	44.76 $\rightarrow$ 45.96 ( +1.20 )	57.15 $\rightarrow$ 55.60 (-1.55)
Dome-DETR-M	30.86 $\rightarrow$ 33.02 ( +2.16 )	50.80 $\rightarrow$ 51.10 ( +0.30 )	64.86 $\rightarrow$ 65.13 ( +0.27 )	32.02 $\rightarrow$ 33.46 ( +1.44 )	48.46 $\rightarrow$ 49.41 ( +0.95 )	60.76 $\rightarrow$ 59.80 (-0.96)
Dome-DETR-L	31.54 $\rightarrow$ 32.90 ( +1.36 )	50.12 $\rightarrow$ 52.09 ( +1.97 )	68.34 $\rightarrow$ 68.86 ( +0.52 )	31.71 $\rightarrow$ 33.37 ( +1.66 )	49.15 $\rightarrow$ 50.40 ( +1.25 )	55.60 $\rightarrow$ 56.62 ( +1.02 )

Appendix

Table 13: Density-wise analysis on the VisDrone validation set. Images are ranked by the number of annotated instances and divided into two equally sized sparse and dense subsets. Each entry reports baseline

\rightarrow

ReDHF, with the absolute change in parentheses.

Assignment strategy	AP	AP 50	AP 75	AP vt	AP t	AP s	AP m
Rectangle-based	33.2	66.9	28.7	17.6	33.0	38.2	45.0
Short-axis circle-based	33.3	67.8	28.4	18.1	32.9	38.4	46.1
Ellipse-guided (ours)	33.5	68.1	28.9	18.9	33.5	38.7	45.8

Appendix

Table 14: Comparison of different label assignment strategies for EGSS. The best result in each column is highlighted in bold.

Inject layer	AP	AP 50	AP 75	AP vt	AP t	AP s	AP m
$P_{2}\sim P_{3}$	32.8	66.5	28.1	18.0	32.7	38.5	45.0
$P_{2}\sim P_{5}$	32.4	66.3	27.3	18.3	32.2	38.2	44.6
$P_{2}$ only	33.5	68.1	28.9	18.9	33.5	38.7	45.8

Appendix

Table 15: Ablation on the detail injection level. We keep other components unchanged and only vary the target feature levels. The best result in each column is highlighted in bold.

Aggregation mode	Params	GFLOPs	AP	AP 50	AP 75	AP vt	AP t	AP s	AP m
Bilinear	15.88	182.3	29.4	61.9	24.0	14.9	29.4	36.6	41.5
AvgPool	15.88	182.3	33.1	66.9	28.7	18.4	32.9	38.8	45.0
Strided Conv	15.95	187.5	32.8	66.7	27.9	18.1	32.8	38.0	44.7
Deform. Cross-Attn	15.93	188.2	33.5	68.1	28.9	18.9	33.5	38.7	45.8

Appendix

Table 16: Comparison of different strategies for downsampling

P_{1}

to the

P_{2}

resolution in CSDF. We keep the semantically enhanced

P_{1}

and other ReDHF components unchanged, and only vary the strategies. The best result in each column is highlighted in bold.

Aggregation	$\mu_{\mathrm{in}}(P_{2}^{\prime})$	$\mu_{\mathrm{out}}(P_{2}^{\prime})$	FBR $(P_{2}^{\prime})$	FBR $(P_{2})$	$\Delta$ FBR	AP vt
Bilinear	0.49	0.33	1.48	1.44	0.04	14.9
AvgPool	0.45	0.27	1.67	1.45	0.22	18.4
Strided Conv	0.52	0.32	1.63	1.45	0.18	18.1
Deformable cross-attn	0.57	0.31	1.84	1.46	0.38	18.9

Appendix

Table 17: Foreground-to-background response analysis for different downsampling strategies in CSDF.

\Delta

FBR denotes the increase from

P_{2}

P_{2}^{\prime}

. The best result in each column is highlighted in bold.

Figure 5: Sensitivity analysis of the sampling-point numbers in CSDF and FSQI. The bottom x-axis shows the number of sampling points, the top x-axis shows the corresponding GFLOPs. The left y-axis reports AP, while the right y-axis reports $AP_{vt}$ . In the left subplot, we vary the CSDF points while fixing the FSQI points to 8. In the right subplot, we vary the FSQI points while fixing the CSDF points to 8.

Dataset	Baseline AP / Tiny AP	GT coverage @8 / @16 (%)	AP / Tiny AP @8	AP / Tiny AP @16	Default $r_{\max}$
AI-TOD-v2	32.1 / 17.5	98.17 / 100.00	33.5 / 18.9	33.4 / 18.7	8
VisDrone	35.9 / 28.3	64.14 / 90.28	37.1 / 29.8	37.5 / 30.3	16
SODA-D	35.4 / 17.6	73.69 / 100.00	36.6 / 19.3	36.8 / 19.6	16

Appendix

Table 18: Sensitivity to the scale threshold

r_{\max}

in EGSS. GT coverage denotes the percentage of valid ground-truth instances selected for supervising

P_{2}^{\prime}

under the corresponding threshold. Tiny AP denotes AP vt , AP s , and AP eS on AI-TOD-v2, VisDrone, and SODA-D, respectively. The adopted setting for each dataset is highlighted in bold.

Setting	$\lambda_{\text{cls}}$	$\lambda_{\text{reg}}$	AP	AP 50	AP vt	AP t	AP s
cls only	1.0	–	33.2	68.2	18.1	33.2	38.5
cls + reg	1.0	0.10	33.5	68.1	18.9	33.5	38.7
cls + reg + ctr	1.0	0.10	32.1	66.5	15.8	32.2	38.2
cls + reg	1.0	0.05	33.3	67.1	18.7	33.3	39.2
	1.0	0.20	33.1	67.2	17.6	33.7	38.6
	1.0	0.50	32.8	67.7	17.3	32.3	38.6

Appendix

Table 19: Ablation study on the EGSS loss design. The centerness loss weight is set to 0.1 when used. The best result is highlighted in bold.

Table 23

Scale	Num. GT	Avg. $N_{\mathrm{pos}}$	$N_{\mathrm{pos}}\leq 2$ (%)	Avg. $N_{c\geq 0.5}$	$N_{c\geq 0.5}\leq 1$ (%)
Very tiny	47317	1.90	78.26	0.54	93.90
Tiny	276865	6.44	23.69	1.79	33.76
Small	45612	19.19	3.18	5.21	1.57
Medium	6782	81.05	0.13	21.59	0.93

Appendix

Table 22: Statistics of centerness targets across object scales on AI-TOD-v2.

N_{\mathrm{pos}}

denotes the number of positive feature locations assigned to each ground-truth instance, and

N_{c\geq 0.5}

denotes the number of positive locations whose centerness target is at least 0.5.

Small Object Detection

Small object detection develops methods for finding and localizing objects that occupy few pixels, often amid clutter, blur, or low resolution.

See all 91 papers

FSDETR: Frequency-Spatial Feature Enhancement for Small Object Detection
Apr 16, 2026Jianchao Huang, Fengming Zhang, Haibo Zhu +1Frequency-Domain Feature Learning Small Object Detection
School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, China
TinyFormer: Preserving Tiny Objects in YOLO-DETR Hybrid Real-time Detectors
May 24, 2026Jun-Wei Hsieh, Meng-Yu Kao, Ghufron Wahyu Kurniawan +1Hybrid CNN-Transformer Architectures Small Object Detection
College of Artificial Intelligence, National Yang Ming Chiao Tung University, Taiwan · 2Mitsubishi Electric Research Laboratories
UHR-DETR: Efficient End-to-End Small Object Detection for Ultra-High-Resolution Remote Sensing Imagery
Apr 23, 2026Jingfang Li, Haoran Zhu, Wen Yang +4Remote Sensing Small Object Detection Small Object Detection
School of Electronic Information, Wuhan University, Wuhan, 430072 China

Apr 16, 2026J/K move · Enter open · S save

FSDETR: Frequency-Spatial Feature Enhancement for Small Object Detection

Jianchao Huang, Fengming Zhang, Haibo Zhu, Tao Yan

School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, China

Frequency-Domain Feature Learning Small Object Detection Detection Transformer Multi-Scale Feature Fusion Aerial Object Detection

Open paper AskPDF arXiv

Small object detection remains a significant challenge due to feature degradation from downsampling, mutual occlusion in dense clusters, and complex background interference. To address these issues, this paper proposes FSDETR, a frequency-spatial feature enhancement framework built upon the RT-DETR baseline. By establishing a collaborative modeling mechanism, the method effectively leverages complementary structural information. Specifically, a Spatial Hierarchical Attention Block (SHAB) captures both local details and global dependencies to strengthen semantic representation. Furthermore, to mitigate occlusion in dense scenes, the Deformable Attention-based Intra-scale Feature Interaction (DA-AIFI) focuses on informative regions via dynamic sampling. Finally, the Frequency-Spatial Feature Pyramid Network (FSFPN) integrates frequency filtering with spatial edge extraction via the Cross-domain Frequency-Spatial Block (CFSB) to preserve fine-grained details. Experimental results show that with only 14.7M parameters, FSDETR achieves 13.9% APS on VisDrone 2019 and 48.95% AP50 tiny on TinyPerson, showing strong performance on small-object benchmarks. The code and models are available at https://github.com/YT3DVision/FSDETR.

Abstract

Small Object Detection

Explore similar work

FSDETR: Frequency-Spatial Feature Enhancement for Small Object Detection

TinyFormer: Preserving Tiny Objects in YOLO-DETR Hybrid Real-time Detectors

UHR-DETR: Efficient End-to-End Small Object Detection for Ultra-High-Resolution Remote Sensing Imagery