TERA: Task-Evidence Restoration for Underwater Object Detection under Bandwidth-Constrained Transmission

Heejin Kim, Joon-Seok Song, Eunil ParkarXiv 2026 · arXiv

TL;DR

TERA is a new framework designed to rescue underwater object detection from the "digital murky waters" of low-bandwidth transmission. By combining task-guided pixel restoration with a novel dual-state feature correction, TERA significantly boosts detection accuracy on compressed and downscaled images without requiring any changes to the original, pre-trained detector.

What Problem Does This Paper Address?

Underwater robots are vital for marine research and infrastructure inspection, but they face a double-edged sword when it comes to vision. First, they deal with physical optical degradation like color distortion and turbidity. Second, because underwater communication bandwidth is extremely limited, images must be aggressively downscaled and compressed before being sent to a receiver.

Standard object detectors are usually trained on high-quality, pre-transmission images. When these detectors encounter the "blocky" or "blurry" artifacts of low-bitrate transmission, their performance plummets. Existing solutions often try to fix the image for human eyes (perceptual restoration) or retrain the detector on degraded data. However, retraining can be costly and often degrades performance on high-quality "clean" images. The authors identify a critical gap: a lack of benchmarks for this specific "transmission-induced" degradation and a need for restoration methods that focus on what the detector needs to see, not just what looks good to a human.

Figure 1. Detection under bandwidth-constrained transmission. (a) Pre-transmission image with ground truth. (b) Transmitted im- age after downscaling and compression. (c) Missed detections by the frozen detector. (d) TERA restores task evidence and recovers the missed detections without updating the detector.

Methods

The authors propose TERA (Task-Evidence Restoration), a receiver-side framework that operates while keeping the main object detector completely frozen. It consists of two primary stages:

  1. Pixel Restoration: An encoder-decoder network takes the degraded image and predicts a "gated RGB residual" to clean it up. Crucially, this stage is guided by intermediate features from the frozen detector's backbone, ensuring the restoration focuses on regions important for the task.
  2. Dual-State Evidence Restoration: Even a "restored" image might have subtle feature discrepancies that confuse a detector. TERA addresses this by looking at the detector's internal features for both the degraded image and the pixel-restored image. By comparing these "dual states," the system predicts a correction to the detector's late-stage features ($N_3, N_4$ pyramid levels) to recover residual evidence that pixel-level cleaning might have missed.

To train TERA, the authors use Foreground-restricted teacher distillation. They use the detector's response to the original, clean image as a "teacher" to guide the restorers, focusing specifically on object-related regions rather than the background.

Figure 2. The Overall TERA Pipeline. It uses a frozen detector and adds trainable pixel and evidence restorers to refine both the input image and the internal feature representations.

Datasets

The paper introduces AquaTX, a controlled benchmark for evaluating transmission-aware underwater detection. It is built upon two existing real-world datasets:

  • DUO-TX: Based on the DUO dataset (4 classes: holothurian, echinus, scallop, starfish).
  • RUOD-TX: Based on a 4-class subset of the RUOD dataset.

AquaTX applies systematic resizing and compression using various codecs (JPEG, WebP, JPEG2000) at multiple bitrates. Importantly, it includes a "held-out" codec (AVIF) that is never seen during training, allowing researchers to test how well their models generalize to new compression standards.

Key Insights and Ideas

  • Task-Guided > Perceptual: Restoring an image to look "pretty" for humans isn't the same as restoring it for a machine. TERA's use of detector features to guide restoration is a key differentiator.
  • The "Frozen" Advantage: By keeping the detector frozen, TERA avoids the "catastrophic forgetting" often seen in fine-tuning. It maintains—and even slightly improves—performance on high-quality clean images while gaining robustness to degradation.
  • Dual-State Comparison: The idea that the degraded image might still contain "complementary evidence" that restoration accidentally removes is powerful. By looking at both states, the evidence restorer can make more informed corrections.
  • Metadata-Blind: TERA doesn't need to know the compression rate or codec used at the transmitter; it figures out how to fix the image purely from the received pixels.

Key Results

TERA demonstrates substantial performance gains across the board:

  • Accuracy Boost: Under compound degradation (resizing + compression), TERA improved mAP$_{50:95}$ from 37.7% to 50.0% on DUO-TX and from 31.5% to 41.3% on RUOD-TX compared to a baseline frozen detector.
  • Outperforming Fine-tuning: TERA outperformed a baseline that was specifically fine-tuned on degraded data (DegFT-YOLO) by nearly 2 points on DUO-TX, while also being significantly better on clean images (+0.7 mAP vs -4.3 mAP for fine-tuning).
  • Generalization: TERA remained highly effective on the unseen AVIF codec, showing it learns general principles of restoration rather than just memorizing specific codec artifacts.
  • Efficiency: The additional feature-level restoration stage adds only about 2.5ms of latency on an NVIDIA H200, making it a viable trade-off for the massive accuracy gains.

Limitations and Future Work

The authors acknowledge several areas for future exploration:

  • Architecture Scope: The current study focused on a single detector family (YOLOv8). Testing generalizability across other architectures (like Transformers) is a logical next step.
  • Channel Effects: AquaTX models digital compression and resizing but does not yet account for lower-level transmission issues like packet loss or bit errors.
  • Compute Constraints: While the overhead is relatively low, it might still be a hurdle for the most resource-constrained underwater edge devices. Future work could focus on making the restorers even more lightweight.

Summary updated

Abstract

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.
BibTeX
@misc{kim2026tera,
  author = {Heejin Kim and Joon-Seok Song and Eunil Park},
  title = {{TERA: Task-Evidence Restoration for Underwater Object Detection under Bandwidth-Constrained Transmission}},
  year = {2026},
  eprint = {2611.02419},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.02419},
}
Figure