Telling Compression Damage from Bad Labels in Satellite Object Detection

Satellites collect far more imagery than their radio links can send to the ground. One promising fix runs the first few layers of a detector on the satellite, compresses those intermediate features, and finishes detection on the ground. The ground detector is then trained with modern box-regression losses. The best of these losses deliberately ignore boxes that are badly wrong, on the theory that such boxes come from occlusion, blur, or annotation errors and are not worth learning from. Compression breaks that theory. After aggressive compression, the boxes that are badly wrong are disproportionately tiny objects whose fine detail the compressor threw away. These objects are perfectly learnable if the satellite spends a few more bits on them. The loss silences exactly the examples that should tell the system where its bits are going wrong.

We propose to split each training example's "how wrong is this box" signal into two parts. The first is intrinsic difficulty, measured on the uncompressed image, which is always available at training time on the ground. The second is compression-induced difficulty, the extra error that appears only after the image passes through the compressed channel. The loss keeps suppressing intrinsically hopeless boxes, as current methods do. It stops suppressing compression-damaged ones, and it routes their excess error back into the onboard compressor's bit-allocation objective. The whole mechanism lives in training: the flight software gets a better encoder but no extra onboard compute.

Core Claim

The relevant loss is Wise-IoU v3, from Wise-IoU. It scores each predicted box by an "outlier degree" $\beta$: that box's overlap loss divided by a running average of overlap losses. It then gives the most gradient to boxes of moderate $\beta$ and shrinks the gradient for both easy boxes and extreme ones. This loss improved COCO $AP_{75}$ from 53.03% to 54.50% on YOLOv7-w6. Aerial detectors adopted it for small objects: SSCW-YOLO and MFDA-YOLO did so on drone imagery, and the SAR ship detector with Wise IoU did so on radar imagery. All of these papers compute $\beta$ on pristine inputs.

The onboard compression line, represented by FOOL, makes a different choice. FOOL compresses shallow features onboard and reports a 51% lower bitrate than prior split-compression methods at under 1% mAP loss on DOTA-2, SpaceNet-3 and xView. It trains its ground detector with standard losses. It also names high-frequency loss on tiny objects as an open problem.

Our claim is specific. Suppose the Wise-IoU rule is applied to compressed features. Then the set of boxes it suppresses shifts toward small objects that are recoverable on clean input, and this shift grows as bitrate falls. Separating the two sources of error fixes the loss. Using the compression-induced part as a spatial bit-allocation signal improves the rate-versus-accuracy curve on small objects, at the low bitrates where downlink limits actually bind.

Mechanism

For each matched anchor $i$, the ground trainer runs the detector twice. One pass goes through the full compress-and-decode path and gives overlap loss $L_i^{c}$. A second pass, cheap and gradient-free, uses a frozen teacher on the uncompressed image and gives $L_i^{0}$. Split-compression training already carries such a clean teacher for feature distillation. We define an intrinsic outlier degree $\beta_i^{\text{int}} = L_i^{0}/\overline{L^{0}}$ and a channel excess $\delta_i = \max(0, L_i^{c} - L_i^{0})/\overline{L^{0}}$, where bars denote running means. The regression weight becomes

$$ w_i = r(\beta_i^{\text{int}})\,\big(1 + \lambda\,\delta_i\big), $$

Here $r(\cdot)$ is Wise-IoU v3's unchanged non-monotonic gain function and $\lambda$ is a single scalar. A box that is hopeless even on clean input is still suppressed. A box that is fine on clean input but broken after compression is no longer mistaken for noise.

We then splat the per-anchor excess $\delta_i$ onto the feature grid as a map $D(x)$, and add it to the encoder's rate-distortion objective as a spatial weight on feature distortion:

$$ \mathcal{L}_{\text{enc}} = R + \gamma \sum_x \big(1 + \mu D(x)\big)\,\lVert f(x) - \hat f(x)\rVert^2 . $$

In this objective, $R$ is the estimated bitrate, $f$ and $\hat f$ are the teacher and decoded features, and $\gamma, \mu$ trade rate against weighted distortion. The encoder learns to spend bits where the detector's localization is being destroyed by the channel, rather than wherever the teacher's features merely differ.

This is where the spatial-resolution literature enters, as the failure mode rather than as a component. LIP shows that learned, content-aware downsampling protects small objects: COCO $AP_s$ went from 21.9% to 24.0%. The global-local detectors (multi-branch building detection, GLAD) show that uniformly shrunk images lose tiny targets. We treat learned pooling in the onboard encoder as an optional ablation, not the contribution. It is a competing way to protect small objects, and it is trained by ordinary task gradients with no notion of which errors the channel caused.

Research Program

The work has three stages.

The first is a diagnostic study that could kill the idea early. We will reproduce a FOOL-style split encoder at several bitrates on xView and DOTA-2 and train ground detectors with plain CIoU and with Wise-IoU v3. For every suppressed anchor we will log its clean-input $\beta^{\text{int}}$, its object size, and its $\delta$. The central measurement is the fraction of Wise-IoU-suppressed anchors that sit in the moderate, learnable band on clean input, plotted against bitrate. If this fraction stays near its uncompressed value, the tension described above does not exist, and we stop and report that.

The second stage trains the decomposed loss alone, with the encoder frozen. This isolates the loss effect.

The third stage adds the $D(x)$ feedback to encoder training and measures full rate-accuracy curves. We will also check whether the running-mean denominators stay stable at the small batch sizes typical of this setup, a known fragility of Wise-IoU. If they do not, we will replace the running means with per-image medians.

Baselines

The naive reference sends pixels through a conventional codec (JPEG2000 at matched bitrate) or downsamples the image uniformly to meet the budget, then runs the same ground detector with CIoU. xView already shows how badly fixed small chips do: 0.146 mAP against 0.259 for multi-resolution SSD. This baseline anchors how much any learned scheme buys.

The strongest existing method is FOOL itself. It is the only system in the file that addresses this exact downlink-plus-detection setting and runs in real time on a 15W Jetson Orin Nano. Its reported result is a 51% lower bitrate than prior shallow variational bottleneck methods at under 1% mAP drop on DOTA-2, SpaceNet-3 and xView. The file gives no absolute mAP or small-object breakdown for FOOL, so we will re-run it ourselves. We will use the same backbone split, training data, number of epochs, and hyperparameter search budget as our method, and sweep it over the same bitrate grid. Beating it means one of two things. Either we achieve higher $AP_s$ at equal bitrate in the low-rate regime, with overall mAP no worse. Or we reach FOOL's mAP at a measurably lower bitrate, reported as average bitrate saving across the curve.

The second strong baseline is FOOL's encoder with Wise-IoU v3 applied naively to the ground detector. This is the "obvious" combination the file flags as untried, and our diagnosis predicts it hurts small objects at low rate. On uncompressed drone imagery, Wise-IoU-based detectors report 37.8% mAP50 on VisDrone2019 (SSCW-YOLO). This calibrates the uncompressed ceiling for our small-object stress test.

Ablations separate the ingredients. The first uses $\beta^{\text{int}}$ only, with no excess term. The second adds the excess term to the loss without encoder feedback. The third uses encoder feedback driven by teacher feature error alone, without detection-derived $\delta$. The fourth adds LIP-style learned pooling in the onboard encoder. All runs share data, compute, and tuning budget.

Why not just…?

Why not just drop dynamic focusing and use CIoU on compressed features? It is possible that suppression is simply harmful under compression and plain monotonic losses are enough. Wise-IoU's gains on uncompressed data (Wise-IoU) and on noisy SAR imagery (Enhanced Shuffle Attention SAR) suggest that suppressing genuinely bad labels still matters, because satellite datasets contain annotation noise. Our method keeps that benefit and removes the collateral damage. The CIoU baseline is run at every bitrate. If it matches our loss-only variant, the loss half of the contribution is void, and only the encoder feedback remains to be judged.

Why not just train the encoder end-to-end with the detection loss? End-to-end task gradients already push bits toward useful regions, and this is the most serious alternative. The difference is that the detection gradient reaching the encoder is itself filtered by whatever box loss is used. With Wise-IoU, compression-damaged small objects contribute almost nothing, so end-to-end training inherits the blind spot. We will include an end-to-end variant, trained with CIoU and with Wise-IoU, as an extra baseline. If CIoU end-to-end matches our $D(x)$ feedback, we will say that the decomposition adds nothing beyond "don't suppress."

Why not just upweight small objects by box size? The compression-induced excess may largely track object size, and a size prior is free. This is a real threat. The diagnostic study measures how much of the variance in $\delta$ is explained by object area and local background texture. A size-reweighted Wise-IoU, with its size curve tuned on the same budget, is run as a baseline. If size explains most of the variance and size weighting matches our results, we will report the simpler method as the answer. Our expectation is that $\delta$ also captures medium objects in cluttered or low-contrast tiles that a size prior misses, but the experiment decides.

Proposed Experiments

The main benchmarks are xView and DOTA-2, the satellite and aerial detection sets FOOL evaluates on. VisDrone2019 serves as a tiny-object stress test, and HRSID serves as a SAR check, since speckle is a second source of genuine outliers that our loss must still suppress. Each method is trained over a grid of target bitrates. We report mAP, $AP_s$, rate-accuracy curves, and average bitrate saving at matched mAP, plus onboard latency on a Jetson-class device to confirm that the flight-side cost is unchanged.

Three results would be decisive. First, the diagnostic curve must show recoverable small objects making up a growing share of suppressed anchors as bitrate drops. Second, at the lowest bitrates the decomposed loss must beat naive Wise-IoU on $AP_s$, with gains larger than those from size reweighting. Third, encoder feedback must move the full rate-accuracy curve past our re-run of FOOL in the low-rate regime.

If the first result fails, the premise is wrong. If the second succeeds but the third does not, the contribution shrinks to a better ground-side loss for compressed features, which is still worth reporting honestly.