Source-linked AI summary
CSCWD: Cross-Scale Channel-wise Knowledge Distillation for Lightweight Tiny Object Detection on Edge Devices
Amir Zamani, Zeinab Ghasemi-Naraghi
TL;DR
Tiny objects offer weak spatial evidence, and lightweight detectors can lose high-resolution detail. CSCWD transfers teacher features across scales during training while retaining the student inference architecture; it improves Drone-vs-Bird mAP@0.5 over CA-YOLO11n and same-scale CWD.
Problem
Tiny objects provide limited appearance information and weak localization evidence, while lightweight detectors can lose the high-resolution cues needed for their localization.
Source: §1 · p. 1Method
CSCWD transfers high-resolution features from a YOLO11m-P2 teacher to a YOLO11n student during training while leaving the student's inference graph unchanged.
Source: §1 · p. 2Results
50.17% mAP@0.5 under the unified seven-sequence Drone-vs-Bird protocol improves CA-YOLO11n by 2.92 points, and cross-scale alignment adds 2.09 points over same-scale CWD.
Source: §6 · p. 11Takeaways & Limitations
CSCWD improves tiny-target detection in aerial UAV imagery while preserving the deployed YOLO11n inference graph.
Source: §1 · p. 2, §6 · p. 11Evaluation remains concentrated on aerial imagery, and the zero-shot result does not establish broad generalization across unseen tiny-object settings.
Source: §5.3 · p. 10
Abstract
from arXiv · showhide
Real-time tiny object detection in aerial imagery is constrained by the weak spatial evidence of very small objects and the loss of high-resolution detail in lightweight detectors. This study presents Cross-Scale Channel-wise Knowledge Distillation (CSCWD), a training-time framework that transfers high-resolution spatial representations from a YOLO11m-P2 teacher to a compact YOLO11n student without altering the student's inference architecture. Unlike conventional same-scale feature distillation, CSCWD transfers supervision from teacher P2 to student P3 after feature alignment while retaining same-scale distillation at deeper pyramid levels. Under the unified seven-sequence Drone-vs-Bird validation protocol, YOLO11n-CSCWD achieves 50.17% mean average precision at an intersection-over-union threshold of 0.5 (mAP@0.5) and 59.73% recall, improving the matched CA-YOLO11n baseline by 2.92 percentage points in mAP@0.5 and 3.55 points in recall. Cross-scale alignment further increases mAP@0.5 by 2.09 points over the corresponding same-scale channel-wise distillation configuration. In zero-shot evaluation on DUT-Anti-UAV, mAP@0.5 increases from 48.29% to 50.06% without target-domain fine-tuning. This domain was included because its challenging small targets make low-latency, computationally efficient detection particularly relevant. On Raspberry Pi 5 using NCNN-FP16 at 640x640 resolution, the 2.58-million-parameter student achieves 50.32% mAP@0.5 at 82.32 ms mean wall-clock latency, or 12.15 frames per second, while retaining essentially the same runtime and memory requirements as the matched baseline. The results support cross-scale distillation for improving tiny-target detection without increasing inference-time model complexity.
1 Introduction
Tiny targets provide weak spatial and localization evidence, while lightweight detectors may lose the high-resolution detail needed to detect them. CSCWD transfers teacher features across scales during training, and the study evaluates its accuracy and deployment speed on aerial imagery.
1 Introduction
- Tiny targets spanning only a few pixels provide limited appearance information and weak localization evidence, compounded in long-range UAV detection by blur, low contrast, and clutter.Aerial detectors also face latency, memory, and compute constraints on embedded hardware.Source: §1 · pp. 1–2
- Lightweight detectors reduce inference complexity but can lose shallow high-resolution cues, motivating training-time enrichment that does not enlarge the network.Prior multi-scale, higher-capacity, and tracking-assisted strategies can add computation and memory demand.Source: §1 · p. 2
- CSCWD transfers high-resolution teacher P2 supervision to student P3 while retaining same-scale distillation at P4 and P5, then removes teacher components so the student inference graph is unchanged.The framework uses a YOLO11m-P2 teacher and a standard YOLO11n student.Source: §1 · p. 2
- The study evaluates CSCWD against CA-YOLO11n and lightweight detectors on Raspberry Pi 5, with Fig. 1 summarizing deployed accuracy and inference speed.The figure uses NCNN-FP16 at 640×640 under the unified seven-sequence validation protocol.Source: §1 · p. 2, Figure 1 · p. 2
2 Related Work
Prior tiny-object methods strengthen weak-target representation through multi-scale processing, added capacity, or temporal information, but can increase deployment complexity. The related work motivates training-time transfer that supplies teacher-only high-resolution features without changing the student inference graph.
2 Related Work
- Drone-vs-Bird approaches use multi-scale processing, WideResNet, or tracking-assisted recovery, while added computation or pipeline complexity can limit low-complexity deployment.Laroca et al. combine multi-scale processing with augmentation and temporal post-processing; Wong et al. use WideResNet; Sawada et al. use tracking.Source: §2 · pp. 2–3
- Training-only context-aware augmentation improves a nanoscale detector without modifying its inference graph, but does not provide teacher-derived feature supervision.Source: §2 · p. 3
- Feature-level distillation methods transfer representations during training, but compatible-scale matching leaves no counterpart for teacher P2 when the lightweight student begins at P3.ScaleKD addresses scale-aware small-object distillation with a scaledecoupled feature module and cross-scale assistant. CWD represents feature knowledge as channel-wise spatial probability distributions.Source: §2 · p. 3
3 Proposed Method
CSCWD transfers teacher-only high-resolution representations to a lightweight student during training while preserving the student's original inference architecture.
3 Proposed Method
- CSCWD transfers teacher-only high-resolution representations to a lightweight student during training while preserving the student's original inference architecture.Source: §3 · p. 3
- The distillation framework is training-time, and the deployed model retains the original student inference architecture.Source: §3 · p. 3
- The method's core design pairs teacher representations unavailable in the lightweight student's native feature hierarchy with that student during training.Source: §3 · p. 3
3.1 Framework Overview
The framework uses an asymmetric teacher–student setup to transfer high-resolution supervision unavailable in the student's native pyramid. Only the YOLO11n student is retained for deployment.
3.1 Framework Overview
- The frozen YOLO11m-P2 teacher supplies cross-scale P2-to-P3 supervision and matched-scale distillation at P4 and P5, while YOLO11n is retained for deployment.Source: §3.1 · p. 3
- Teacher-only high-resolution supervision is transferred to student P3 during training, alongside same-scale supervision at the deeper pyramid levels.Source: §3.1 · p. 3
- The YOLO11n student remains the only network retained for deployment.Source: §3.1 · p. 3
3.2 Asymmetric Teacher–Student Design
The design pairs a higher-capacity YOLO11m-P2 teacher with a lightweight YOLO11n student. The teacher provides higher-resolution spatial information during training while the student retains its standard architecture.
3.2 Asymmetric Teacher–Student Design
- The YOLO11m-P2 teacher adds a high-resolution P2 detection branch that produces a 160×160 feature map for a 640×640 input.The teacher is trained through VisDrone adaptation followed by target-domain fine-tuning.Source: §3.2 · p. 3
- The YOLO11n student has no P2 branch and begins its detection pyramid at P3.It is initialized from COCO-pretrained weights and adapted on VisDrone before CSCWD training.Source: §3.2 · p. 3
- The teacher–student asymmetry exposes the student to spatial information available only at the teacher’s higher-resolution P2 level while preserving the lightweight student graph.Source: §3.2 · p. 3
3.3 Cross-Scale Feature Alignment
CSCWD aligns teacher P2 with student P3 for cross-scale supervision, while P4 and P5 remain scale matched. The section also identifies teacher training settings and summarizes the training-to-inference flow.
3.3 Cross-Scale Feature Alignment
- Teacher P2 is paired with student P3, while P4 and P5 are distilled at matched scales.Fig. 2 shows that the frozen teacher supplies these signals during training and only the distilled YOLO11n-CSCWD remains at inference.Source: §3.3 · p. 3, Figure 2 · p. 4
- Table 1 summarizes teacher training settings for VisDrone adaptation and target fine-tuning.Source: Table 1 · p. 3
- For the P2-to-P3 pair, the student feature is upsampled to the teacher’s spatial resolution and projected to its channel dimension; P4 and P5 require channel projection only.Source: §3.3 · p. 4
3.4 Cross-Scale Channel-wise Knowledge Distillation
CSCWD extends channel-wise distillation to cross-scale feature pairs by aligning student features to teacher features and comparing their spatial distributions. The training objective combines detection and distillation losses, while deployment retains only the distilled student.
3.4 Cross-Scale Channel-wise Knowledge Distillation
- CSCWD extends channel-wise distillation to compare teacher and student channel-wise spatial distributions using KL divergence, including cross-scale feature matching.The feature pairs are P3→P2, P4→P4, and P5→P5.Source: §3.4 · p. 4
Channel alignment and spatial matching
- A learnable 1 × 1 convolution projects student features into the teacher feature space, and bilinear interpolation matches spatial resolution when needed.After alignment, teacher and student features share identical dimensions for comparison.Source: Channel alignment and spatial matching · p. 5
Channel-wise probability distributions
- Each channel is independently normalized over spatial positions with a temperature-scaled softmax to form a spatial probability distribution.This normalization lets the student learn the teacher’s spatial attention patterns.Source: Channel-wise probability distributions · p. 5
Channel-wise distillation loss
- The distillation loss minimizes KL divergence between teacher and aligned student channel distributions and aggregates the feature mappings.Ck denotes the number of channels at the aligned feature level, and αCWD controls the distillation term’s contribution.Source: Channel-wise distillation loss · p. 5
Total training objective
- The total training objective combines the original detection loss with the proposed distillation loss.Source: Total training objective · p. 5
- After training, teacher, projection, and spatial-alignment components are discarded, leaving the distilled YOLO11n student with its original inference graph and computational footprint.Source: Total training objective · p. 5
4 Experiments and Results
Experiments evaluate CSCWD on Drone-vs-Bird, across sequences and in zero-shot DUT-Anti-UAV testing, then isolate distillation choices and measure Raspberry Pi 5 deployment. The results assess detection gains alongside inference cost and operating conditions.
4.1 Experimental Setup and Evaluation Protocol
- The seven-sequence validation set contains 12,381 labeled frames, with sequence metrics averaged using equal weights; the remaining 70 of 77 sequences are used for target-domain training.Zero-shot DUT-Anti-UAV evaluation uses 2,200 test images without target-domain fine-tuning.Source: §4.1 · pp. 5–6
4.2 Benchmark Comparison on Drone-vs-Bird
- YOLO11n-CSCWD improves CA-YOLO11n by 2.92 percentage points in mAP@0.5 and exceeds YOLO11s by 4.00 points while retaining the 2.58 M-parameter, 6.4-GFLOP YOLO11n graph.The comparison uses the unified seven-sequence validation protocol; the larger YOLO11m-P2 teacher is used only during training.Source: §4.2 · p. 6, Table 2 · p. 7
4.3 Sequence-Level and Cross-Dataset Analysis
- YOLO11n-CSCWD outperforms CA-YOLO11n on six of seven sequences, but GoPro 002 falls from 28.86% to 23.48% mAP@0.5.The largest sequence gains are on Swarm (+10.34 points) and Hillside (+9.23 points); target scale alone does not explain sequence difficulty.Source: §4.3 · p. 6, Table 3 · p. 7
- In zero-shot DUT-Anti-UAV evaluation, YOLO11n-CSCWD improves CA-YOLO11n by 1.77 points in mAP@0.5, while mAP@0.5:0.95 remains essentially unchanged (24.00% versus 24.04%).The checkpoints are applied directly to the DUT test split without target-domain fine-tuning.Source: §4.1 · p. 6, §4.3 · p. 7, Table 4 · p. 7
4.4.1 Distillation Objective
- YOLO11n-CSCWD achieves the highest Recall (59.73%) and 50.17% mAP@0.5, while MGD has the highest mAP@0.5 (50.53%) and mAP@0.5:0.95 (18.00%) among evaluated objectives.The 0.36-point mAP@0.5 difference does not establish universal CWD superiority.Source: §4.4.1 · p. 7, Table 5 · p. 8
4.3 Sequence-Level and Cross-Dataset Analysis
- Fig. 3 presents selected successful tiny-target detections under challenging visual conditions, comparing ground truth with YOLO11m-P2, CA-YOLO11n, and YOLO11n-CSCWD.The sequence analysis notes that target scale alone does not explain difficulty, and GoPro 002 is the clearest CSCWD failure case.Source: §4.3 · p. 6, Figure 3 · p. 8
4.4.2 Teacher Design and Cross-Scale Alignment
- With the P2-enhanced teacher, cross-scale CSCWD reaches 50.17% mAP@0.5 versus 48.08% for same-scale CWD, a 2.09-point gain; Recall rises from 55.36% to 59.73%.The reported ablation distinguishes the cross-scale pairing contribution from teacher capacity alone.Source: §4.4.2 · p. 8
4.5 Embedded Deployment on Raspberry Pi 5
- At 640 × 640 on Raspberry Pi 5, YOLO11n-CSCWD reaches 50.32% mAP@0.5 while both models remain at approximately 82.3 ms mean wall-clock latency and 12.15 FPS.At 960 × 960, the mAP@0.5 advantage is 4.86 points (52.85% versus 47.99%) with essentially unchanged latency and throughput; at 1280, accuracy decreases.Source: §4.5 · p. 8
5 Discussion
The discussion attributes the student’s gains chiefly to cross-scale access to high-resolution teacher features, while showing that spatial selectivity, input resolution, and deployment conditions qualify the results. It also limits the generalization claims to the evaluated settings.
5.1 Cross-Scale Transfer and Feature Selectivity
- 2.09 points higher mAP@0.5 comes from teacher-P2 to student-P3 pairing than same-scale CWD; the larger YOLO11s reference remains 4.00 points lower than YOLO11n-CSCWD.The discussion attributes the gain to cross-scale feature pairing rather than teacher capacity alone; MGD has higher mAP metrics, while CSCWD has the highest Recall.Source: §5 · p. 8, Table 6 · p. 9, §5.1 · p. 9
5.2 Resolution and Edge-Deployment Trade-offs
- At 640 × 640, CSCWD reaches 50.32% mAP@0.5 at 12.15 FPS; at 960 × 960, it reaches 52.85% at 5.45 FPS, before falling to 48.53% at 1280 × 1280.Both detectors were trained at 640, and train–test scale mismatch may contribute to the high-resolution degradation, though the experiment does not isolate that mechanism; fixed-resolution runtime and memory remain essentially identical.Source: Table 7 · p. 9, §5.2 · p. 9, YOLO11m-P2 · p. 10
5.1 Cross-Scale Transfer and Feature Selectivity
- 62.7% of paired frames favor YOLO11n-CSCWD on nTCR (p < 0.001), but the modest, sequence-dependent diagnostic does not establish that concentration improves detection.In the selected Hillside example, CSCWD has a more target-focused P3 response than CA-YOLO11n; GoPro 002 has higher nTCR despite lower sequence-level mAP@0.5.Source: §5.1 · p. 9, Figure 4 · p. 10
YOLO11m-P2
- 11.85 FPS and 72.49°C were recorded for CSCWD versus 0.94 FPS and 78.13°C for YOLO11m-P2; neither run showed throttling.The comparison used one 10-minute sustained-load run per model, so it is descriptive rather than a statistical estimate of long-term thermal reliability.Source: YOLO11m-P2 · p. 10, Figure 5 · p. 10
5.3 Limitations and Future Directions
- Zero-shot DUT-Anti-UAV results support cross-dataset evidence, not broad generalization: evaluation remains concentrated on aerial imagery, with UAVs as representative tiny targets.The discussion also notes that resolution sensitivity uses models trained at 640, nTCR is correlational, and thermal results come from one run per model.Source: §5.3 · p. 10
6 Conclusion
The conclusion presents CSCWD as training-time cross-scale supervision that preserves the YOLO11n inference graph. It reports gains on the seven-sequence Drone-vs-Bird protocol, zero-shot DUT-Anti-UAV, and Raspberry Pi 5 deployment.
6 Conclusion
- Under the unified seven-sequence Drone-vs-Bird validation protocol, YOLO11n-CSCWD achieves 50.17% mAP@0.5, improving CA-YOLO11n by 2.92 percentage points; P2-to-P3 alignment adds 2.09 points over same-scale CWD.Source: §6 · p. 11
- 1.77 mAP@0.5 points are added on zero-shot DUT-Anti-UAV evaluation without target-domain fine-tuning.Source: §6 · p. 11
- 50.32% mAP@0.5 at 12.15 FPS is achieved on Raspberry Pi 5 with NCNN-FP16, with runtime essentially unchanged from CA-YOLO11n.The student retains the YOLO11n inference graph, avoiding teacher-side complexity at deployment.Source: §6 · p. 11
Statements and Declarations
The authors report no funding or competing interests, describe their contributions, and state the availability conditions for data, outputs, logs, and code.
Statements and Declarations
- No funding was received, and the authors report no relevant financial or non-financial competing interests.Source: Statements and Declarations · p. 11
- Amir Zamani is credited with conceptualization, methodology, software, validation, formal analysis, investigation, data curation, visualization, and original drafting; Zeinab Ghasemi-Naraghi with supervision, methodology, project administration, and review and editing.Both authors reviewed and approved the final manuscript.Source: Statements and Declarations · p. 11
- Drone-vs-Bird is available from challenge organizers upon request and subject to access conditions; DUT-Anti-UAV and VisDrone are publicly available, while code and configuration files are not public at preprint release.Derived evaluation outputs and deployment logs are available from the corresponding author upon reasonable request, subject to dataset licensing conditions.Source: Statements and Declarations · p. 11