Bandwidth-limited aerial cameras must compress their images before a remote detector can find the many tiny objects in them. Threshold-based region-of-interest (ROI) coding sends only confidently detected regions, and because tiny objects provide limited visual evidence and often score low, it can discard the regions that contain them. We propose rate-utility region coding (RUO), which sends a region when its expected detection utility justifies its measured byte cost, choosing to drop it or to send it at low or high quality with one multiplier shared across images. RUO trains no model, uses a standard image codec and decodes in a few milliseconds. On all 1,610 VisDrone-DET test-dev images, RUO uses 42.4% fewer bits than a tuned threshold-based ROI coder, averaged over their shared tiny-object average precision (AP) range. RUO also needs fewer bits for tiny objects than the AV1 Image File Format (AVIF), High Efficiency Video Coding (HEVC) and three published learned codecs, and its advantage over threshold-based ROI coding persists when the sender and receiver use different detectors. For AP over all object sizes RUO needs 25.4% more bits than the tuned ROI coder, because its allocation drops large regions first at low rates, so it suits tasks that target tiny objects. We plan to release the code.
Figures & tables
Figure 2: RUO workflow with frozen sender E and receiver D : calibrated utility and measured byte costs guide a shared- λ choice of drop, low or high quality; selected regions are decoded for detection.
Figure 3: Recall decomposition on VisDrone test-dev with sender and receiver D1, showing the outcome of every tiny ground-truth object at each method’s operating point closest to 0.2 bpp, actual rate in brackets (top), and the share of objects of area ≥322 that are selection misses against rate (bottom).
RUO versus
AP tiny BD-rate
AP BD-rate
sender D1
ROI-P. tuned
−42.4
+25.4
ROI-P. τ=0.25
−52.3
+24.1
ROI-P. small only
−19.6
−14.0
AVIF
−77.3
−24.9
HEVC intra, x265
−74.9
−32.1
Table 1: BD-rate (%) of RUO on all 1,610 VisDrone test-dev images , receiver D1 with sliced inference, where a negative value means fewer bits than the baseline for the same AP and bold highlights the tuned ROI comparisons (95% CIs in Tab. S4 of the supplementary material).
Figure 4: Rate–accuracy on VisDrone test-dev for tiny objects (left) and all sizes (right) with sender and receiver D1, each curve the Pareto envelope of a method’s operating points and the dotted line the uncompressed images.
Codec (published weights)
vs AVIF
RUO vs codec
mbt2018-mean [ 31 ]
−28.5
−44.0
cheng2020-anchor [ 8 ]
−35.0
−39.5
TransTIC base [ 7 ]
−22.7
−48.4
Adapt-ICMH [ 26 ]
−20.5
−50.1
SA-ICM [ 34 ] §
–
–
§ One usable rate, AP tiny 9.33 at 0.165 bpp, where RUO reaches 9.82; RUO reaches 9.33 at 0.150 bpp.
Table 2: Learned codecs with published weights, AP tiny BD-rate (%) on the 200 development images with receiver D1, on intervals where no curve has a bpp gap above 2 × , with bold marking the RUO-versus-codec column.
Comparison
AP tiny
AP
VisDrone test-dev, sender D1
small only vs tuned ROI-P.
−27.5 [-33.2, -17.6]
+223.8 [+87.2, +318.2]
same, sender D2
−26.1 [-33.2, -18.6]
+165.1 [+86.8, +277.4]
tuned ROI-P. vs AVIF
−50.5 [-59.4, -27.1]
−34.4 [-41.9, -21.3]
Test-dev, RUO vs tuned ROI-P., BD-AP (points)
sender D1 / D2
+1.17 / +1.20
−1.08 / −0.96
Table 3: Further results , BD-rate (%) for AP tiny and AP over all sizes unless a block states another unit, with 95% CIs over scenes where they were computed.
Figure 5: Qualitative examples illustrating RUO’s tiny-object detection advantage at approximately 0.2 bpp. From left to right: original with tiny ground truth (GT), AVIF, ROI-Packing-lite+ (ROI-P+), and our rate-utility region coding (RUO). Each row is a different image, shown as the fixed 512 × 384-pixel crop. True positives are teal, false positives amber dotted, missed tiny objects coral dashed, and transmitted regions white (dotted for low quality). Pale grayscale context in the last two columns is untransmitted and provides spatial reference only. Actual full-image bpp in AVIF/ROI-P+/RUO order: (a) 0.183/0.207/0.193; (b) 0.185/0.209/0.200; (c) 0.190/0.191/0.207.
Variant
sender D1
sender D2
RUO, fixed tier weight (final)
−28.7
−23.8
RUO-M, measured detectability utility
−20.3
−11.7
raw score instead of calibrated
−17.6
−12.0†
two tiers (drop / high)
−16.5
−7.8
plus low-resolution base layer
−16.0
−6.9
All regions on one canvas, one AVIF image
−8.6
+1.2
Table 4: Ablations : AP tiny BD-rate (%) against tuned ROI-Packing-lite+ on the 200 development images with receiver D1 unless stated, where bold marks the frozen method and indented rows change one component of RUO-M.
Figure 6: Failure cases of region selection at approximately 0.2 bpp, with columns and markings as in Fig. 5 . Actual full-image bpp in AVIF/ROI-P+/RUO order: (a) 0.198/0.219/0.196; (b) 0.200/0.192/0.199.
Method
bpp
enc.
dec.
recv.
GPU MB
RUO (ours)
0.191
340 ∗
7.9
73
688
ROI-Packing-lite+
0.198
175 ∗
8.6
79
688
AVIF
0.159
103
18.1
129
688
JPEG
0.197
18
15.5
170
688
mbt2018-mean
0.165
67
91.9
100
1056
TransTIC base
0.158
172
204.0
99
9573
Table 5: Latency in ms, the median over 21 fixed test-dev images with 5 synchronized runs each, and peak GPU memory on an NVIDIA RTX 6000 Ada Generation, where ∗ marks encoding times that include the sender detector and the receiver runs sliced inference on the decoded image.
Region codec
vs ROI-P.
vs full frame
vs ROI-P., AP
AVIF (main paper)
-28.7
–
+1.2
JPEG
-25.4
-65.6
+28.9
WebP
-8.9
-52.1
+27.5
Table S1 : Region codec , RUO’s BD-rate (%) on the 200 development images with sender and receiver D1 when every region and every baseline uses the given codec, all within the 2 × bpp-gap rule.
Setting
value
AP tiny
AP
frozen
–
-28.7
+1.2
tier weight wlo (0.6)
0.3
-22.8
+42.0
0.9
-15.5
-8.8
low quality qlo (35)
20
-19.2
+6.2
50
-22.8
+42.0
high quality qhi (65)
50
-49.4
+14.5
Table S2 : Sensitivity , BD-rate (%) of RUO with one setting changed against tuned ROI-Packing-lite+ on the development set with sender and receiver D1, with bpp gaps of at most 1.7 × in the AP tiny comparison.
AP tiny at bpp
AP at bpp
Method
0.10
0.20
0.25
0.10
0.20
0.25
RUO (ours)
6.20
7.63
7.76
16.09
20.12
20.95
ROI-P. tuned
4.81
6.63
7.17
18.74
20.67
21.07
ROI-P. τ=0.25
4.54
5.84
6.19
18.54
20.16
20.49
AVIF
2.18
4.58
5.25
13.62
18.34
19.61
JPEG
–
2.19
2.73
–
13.11
14.48
Table S3 : AP tiny and AP (%) at fixed rates on test-dev with sender and receiver D1, where ROI-P. is ROI-Packing-lite+, bold marks the best deployable tiny AP and the uncompressed images reach AP tiny 8.72 and AP 22.92.
RUO versus
AP tiny BD-rate
AP BD-rate
sender D1
ROI-P. tuned
−42.4 [-55.1, -26.6]
+25.4 [+15.9, +37.3]
ROI-P. τ=0.25
−52.3 [-69.4, -38.1]
+24.1 [+12.6, +41.5]
ROI-P. small only
−19.6 [-36.9, +0.2]
−14.0 [-18.4, -6.2]
AVIF
−77.3 [-83.4, -66.2]
−24.9 [-37.7, -4.1]
HEVC intra, x265
−74.9 [-82.4, -60.8]
−32.1 [-45.2, -7.2]
Table S4 : BD-rate (%) of RUO on all 1,610 VisDrone test-dev images with 95% CIs , receiver D1 with sliced inference, where a negative value means fewer bits than the baseline for the same AP, brackets give the CI over 61 scenes and bold highlights the tuned ROI comparisons.
Protocol
Sender
RUO vs
AP tiny
AP
Official AP
Validation calibration,
UAVDT model
ROI-P. tuned
−41.2 [-64.5, -9.0]
+8.8 [-16.8, +44.9]
+27.2
20 test sequences
AVIF
−71.8 [-87.0, -41.7]
−33.9 [-62.7, +20.4]
−34.8
VisDrone D1
ROI-P. tuned
−6.5 [-48.5, +17.8]
+17.1 [-12.8, +67.6]
+43.6
AVIF
−63.6 [-82.6, -22.3]
+5.6 [-33.6, +77.2]
+12.5
Earlier, test calibration,
UAVDT model
ROI-P. tuned
−16.5 [-52.1, +33.7]
+13.3 [-13.6, +52.0]
–
15 test sequences
AVIF
−64.1 [-79.9, -37.7]
−29.4 [-63.4, +10.9]
–
Table S5 : UAVDT , BD-rate (%) of RUO with receiver UAVDT-YOLOv8m and 95% CIs over sequences, under the main paper’s protocol (top, 20 test sequences, 1,667 frames) and the earlier one (bottom, 15 sequences, 1264 frames), with the official toolkit’s class-agnostic AP as a point estimate.
bpp
images
W/L
pROI-P.
pAVIF
RUO
ROI-P.
AVIF
0.1
232
55/68
0.28
0.009
30.9
32.2
31.4
0.2
284
83/78
0.75
0.005
34.7
36.0
32.9
0.3
115
44/29
0.10
0.032
40.2
39.0
38.4
Table S6 : Per-image tiny-object recall on test-dev with sender D1 and each method within 10% of the target rate per image, with W/L the images where RUO’s recall is higher or lower than that of threshold coding (ROI-P.), p the two-sided sign test against ROI-P. and AVIF, and mean recall in %.
#
bpp
tiny
ref.
RUO
strongest
ROI-P.
AVIF
1
0.1
30
10
14
x265
9
7
4
2
0.2
19
15
14
x265
12
10
10
3
0.2
21
9
10
ROI-P.
8
8
4
4
0.1
33
17
16
ROI-P.
13
13
10
5
0.2
23
13
14
ROI-P.
12
12
8
6
0.2
23
11
11
TransTIC
9
1
9
Table S7 : Images where RUO gains most over the strongest of seven baselines at approximately matched per-image rates on test-dev with sender and receiver D1, counting tiny objects found at IoU 0.5 with the class label and score ≥0.05 , with ref. the count on the uncompressed frame and strongest the baseline that finds most tiny objects on that image.
Figure S1 : Ranks 1 and 4 of Tab. S7 at 0.1 bpp, each method at its own point closest to the target on that image, with boxes drawn on the original frame zoomed on the 480×270 window with the most tiny objects, detected tiny objects in teal, false positives with score ≥0.05 in amber dotted and missed tiny objects in coral dashed.
Figure S2 : The teaser image , at the 90th percentile of tiny-object count, with AVIF at 0.190 bpp finding 10/79 tiny objects, threshold coding at 0.191 bpp finding 18/79 and RUO at 0.207 bpp finding 27/79, against 43/79 on the uncompressed frame.
Figure S3 : Median image by tiny-object count, from left to right the original with its tiny ground truth, AVIF, threshold coding (ROI-Packing-lite+) and RUO, each at its point closest to 0.2 bpp on this image.
Figure S4 : 10th-percentile image by tiny-object count, in the layout of Fig. S3 .
Figure S5 : First failure case , an image where no method finds any tiny object at about 0.2 bpp, in the layout of Fig. S3 .
Aerial object detection in UAV imagery presents unique challenges due to the high prevalence of tiny objects, adverse environmental conditions, and strict computational constraints. Standard YOLO-based detectors fail to address these jointly: their minimum detection stride of 8 pixels renders sub-32px objects nearly undetectable, their CIoU loss produces zero gradients for non-overlapping tiny boxes, and their architectures contain significant filter redundancy. We propose DroneScan-YOLO, a holistic system contribution that addresses these limitations through four coordinated design choices: (1) increased input resolution of 1280x1280 to maximize spatial detail for tiny objects, (2) RPA-Block, a dynamic filter pruning mechanism based on lazy cosine-similarity updates with a 10-epoch warm-up period, (3) MSFD, a lightweight P2 detection branch at stride 4 adding only 114,592 parameters (+1.1%), and (4) SAL-NWD, a hybrid loss combining Normalized Wasserstein Distance with size-adaptive CIoU weighting, integrated into YOLOv8's TaskAligned assignment pipeline. Evaluated on VisDrone2019-DET, DroneScan-YOLO achieves 55.3% mAP@50 and 35.6% mAP@50-95, outperforming the YOLOv8s baseline by +16.6 and +12.3 points respectively, improving recall from 0.374 to 0.518, and maintaining 96.7 FPS inference speed with only +4.1% parameters. Gains are most pronounced on tiny object classes: bicycle AP@50 improves from 0.114 to 0.328 (+187%), and awning-tricycle from 0.156 to 0.237 (+52%).
Underwater object detectors are typically evaluated on pre-transmission images, overlooking additional degradation from bandwidth-constrained downscaling and compression. We introduce AquaTX, a controlled benchmark built from real underwater detection datasets with seen and held-out codecs. We also propose TERA, a metadata-blind receiver-side framework combining task-guided pixel restoration and dual-state feature correction while keeping the detector frozen. Under seen compound degradation, TERA improves mean average precision (mAP50:95, averaged over IoU thresholds from 0.50 to 0.95) from 37.7 to 50.0 on DUO-TX and from 31.5 to 41.3 on RUOD-TX relative to the frozen detector without restoration. TERA also preserves clean-input performance and remains competitive with detector fine-tuning under the held-out AVIF codec. These findings support transmission-aware evaluation and task-guided restoration for underwater object detection. Code is available at https://github.com/hijinkim/TERA.
Heejin Kim, Joon-Seok Song, Eunil Park
Sungkyunkwan University, Seoul, Republic of Korea · Jaume I University, Castellon, Spain
Real-time aerial small object detection is an important visual signal and image processing problem, requiring a detector to preserve fine-grained localization while avoiding redundant computation on large background regions. This paper focuses on this deployment-oriented aerial/UAV setting rather than claiming a universal detector for all object detection scenarios. Existing Transformer-based detectors provide strong global modeling, but their dense query initialization and multi-layer decoder still spend substantial computation on background tokens, which is inefficient when small objects occupy only sparse image regions. To address this problem, this paper proposes HGSQ, a Heatmap-Guided Sparse Query Detector for real-time aerial small object detection. HGSQ uses a lightweight Heatmap Budget Predictor (HBP) to predict a foreground budget map in a single forward pass. The predicted heatmap is then used by three fixed components: Heatmap-Guided Sparse Query Selection (HSQS), which initializes decoder queries from high-confidence foreground positions; Heatmap-Gated Lite Snake Convolution (HGLSConv), which performs local shape refinement only on heatmap-activated small-object regions; and Adaptive Query-Decoder Budgeting (AQDB), which adjusts the query budget and decoder depth according to the estimated object density. Unlike post-hoc heatmap generation, HGSQ treats the heatmap as a real-time computation budget rather than a visualization map during deployment. Experiments on NWPU VHR-10 and VisDrone2019 show that HGSQ achieves 95.10 mAP50 on NWPU VHR-10 and 54.8 mAP50 on VisDrone2019, while reducing GFLOPs to 48.6 and running at 96.0 FPS on an RTX 4070 under our TensorRT FP16 deployment protocol.
Yangchen Zeng
School of Cyber Science and Engineering, Southeast University, Nanjing, 211189, China.