FANVIDv2: Evaluating Video Super-Resolution by Face and Licence-Plate Recognition Under Compound Degradation

1 Introduction

• Common VSR benchmarks use bicubic inputs and PSNR/SSIM, while surveillance footage adds blur, noise, and compression that can alter recognition-critical detail.A reconstruction can improve PSNR while changing detail that distinguishes characters such as B from 8.Source: §1 · p. 1
• FANVIDv2 contributes face clips for 48 identities, plate clips, randomized compound degradation, recognition metrics, and a documented baseline family with released scorers.Its baselines report recognition, localisation, and OCR counts, with rates normalized for differing numbers of evaluated frames.Source: §1 · pp. 1–2

2 Related work

• Large VSR models are evaluated almost exclusively with fidelity metrics, whereas FANVIDv2 uses randomized higher-order degradation models for test inputs and baseline training.The cited related work includes REDS, Vimeo-90K, Vid4, BasicVSR++, VRT, and RVRT.Source: §2 · p. 2
• Related benchmarks cover tracklet matching, low-resolution faces, or plates suitable for direct OCR, while FANVIDv2 measures VSR recognition for faces and plates.The two tasks stress different requirements: global structural coherence for faces and character-level precision for plates.Source: §2 · p. 2

3.1 Sources and content

• HR frames are 1280 × 720 and LR frames are 320 × 180; clips come from publicly available YouTube videos and are regenerated using released scripts.Only URLs, timestamps, and annotations are distributed.Source: §3.1 · p. 2
• The face subset covers 48 public figures, each with an HR gallery image from a different source, and training and test identities are disjoint.The gallery image is press photography; clips show public appearances.Source: §3.1 · p. 2
• The plate subset has 375 clips and 360 distinct Latin-alphanumeric strings from more than 25 cities on five continents, split into 260 training and 115 test clips.All Brazilian clips are in test and all Oceanian clips are in training, so test includes plate formats unseen in training; two strings cross splits.Source: §3.1 · p. 2

3.2 Annotation

• Face annotations are manually checked after detection and gallery matching, while plate boxes are propagated from HR key frames and reviewed frame by frame.Plate strings are transcribed manually and normalized to uppercase alphanumerics; face and plate box IoU thresholds are 0.5 and 0.7, respectively.Source: §3.2 · pp. 2–3
• The annotation process reports inter-annotator agreement κ = 0.91, and predictions on distractor faces and plates count as false positives for faces.The supplied table caption identifies dataset composition and says frame counts refer to LR frames of annotated clips.Source: §3.2 · p. 3, Table 1 · p. 3

4 Compound degradation

• Bicubic downsampling alone produces unrealistically clean LR frames, so FANVIDv2 uses one degradation model for benchmark clips and a richer model for VSR training.The models are described as related but serve different purposes.Source: §4 · p. 3

4.1 Benchmark LR generator

• Benchmark LR generation applies randomized blur, resize jitter, noise, and JPEG compression before bicubic ×4 downsampling.Blur and noise are sampled probabilistically; resize jitter uses factors from 0.5 to 1.5, and JPEG quality ranges from 30 to 95.Source: §4.1 · pp. 3–4
• The benchmark degradation parameters are randomized, so test frames span conditions from nearly clean to heavily corrupted rather than one operating point.The degradation pass is sampled independently per frame.Source: §4.1 · p. 4, Figure 2 · p. 4

4.2 Second-order training degradation

• VSR training uses two randomized degradation stages with varied kernels, resampling, noise, and JPEG quality; parameters are drawn once per group of frames.Sharing parameters within a group preserves temporal correlations used by recurrent memory.Source: §4.2 · p. 4, Figure 2 · p. 4

Evaluation flow

• Figure 3 holds the recognisers fixed and varies only the VSR stage, with ground truth annotated on the HR clip.Faces use RetinaFace and ArcFace with gallery cosine similarity; plates use YOLOv10 and EasyOCR without fine-tuning.Source: Figure 3 · p. 5

Face task (T1)

• FaceRecBox credits correct identities against ground-truth boxes plus false positives, so misses and spurious detections both count against the score.The face task matches boxes greedily, one-to-one, in decreasing IoU; a match requires IoU ≥ 0.5.Source: §5 · pp. 4–5

Plate task (T2)

• TextRecBox averages plate box scores over ground-truth plates, combining localisation quality and normalised transcription similarity; missed plates score zero, while false positives are not penalised.The aligned scorer evaluates frames present in both annotations and system output, and maps IoU linearly to a weight from zero at 0.2 to one at 1.Source: §5 · p. 5

6 Reference pipeline and baseline architectures

• The reference pipeline applies off-the-shelf RetinaFace and ArcFace for faces, and YOLOv10 and EasyOCR for plates, without fine-tuning the recognisers.Face identity is assigned by cosine similarity to the gallery.Source: §6 · p. 5

6.1 RCDM

• RCDM takes a group of 2k+1 low-resolution frames (k=3, a seven-frame window) and emits one high-resolution frame, with central-frame Haar-wavelet conditioning producing four half-resolution sub-bands.The sub-bands are upsampled to feature-map resolution, concatenated with features, then linearly projected and added to the residual branch in every ConvNeXt fusion block.Source: §6.1 · pp. 5–6

6.2 Residual-map gated fusion (RCDM-RMGF)

• RCDM-RMGF improves held-out REDS validation PSNR from 22.48 to 22.58 dB and SSIM from 0.591 to 0.598 over pure RCDM under second-order degradation.Its lightweight residual-based gate suppresses the previous-frame memory contribution in regions where consecutive frames differ strongly; the model has 2.35 M parameters.Source: §6.2 · pp. 6–7

6.3 Modular family

• The modular RCDM family varies propagation topology, reconstruction head, block type, and dataset adaptation, with a slim configuration and a plate-specialised variant also described.The plate variant uses glyph-focused patchification and an OCR-coupled objective, but it is not part of the reported results, which use fixed EasyOCR.Source: §6.3 · p. 7

6.4 Baselines

• The baseline comparison evaluates bicubic-upsampled LR, RCDM, and RCDM-RMGF, with both learned systems trained on REDS using second-order degradation and tested on FANVIDv2 without further adaptation.LR is scored after bicubic upsampling; RCDM-RMGF is trained the same way as RCDM.Source: §6.4 · p. 7

Faces

• FaceRecBox rises from 0.6864 to 0.7222 with VSR, while matched-face identity accuracy increases from 84.35% to 86.93%.Mean IoU rises from 0.936 to 0.969, and both precision and recall increase.Source: §7 · p. 7, Table 2 · p. 8

Plates

• TextRecBox rises from 0.3088 to 0.3667 with RCDM and to 0.3801 with RCDM-RMGF on plates.Exact reads per 1,000 ground-truth boxes rise from 22.5 to 28.1 with RCDM; raw counts are confounded by differing evaluated frame counts.Source: §7 · p. 8, Table 3 · p. 8

Fidelity and recognition

• On plate clips, RCDM improves PSNR from 24.26 to 25.97 dB and SSIM from 0.733 to 0.807, but PSNR/SSIM correlate weakly with TextRecBox (Spearman ρ ≈0.3).Their correlation with FaceRecBox is moderate (ρ ≈0.6).Source: §7 · p. 8

What the benchmark measures

• FANVIDv2 measures whether VSR output can be recognised, not whether it looks like the original.Face gains involve localisation and identity matching, while plate gains involve reading characters already detected.Source: §8 · p. 9

Limitations

• The benchmark covers 48 identities and 360 plate strings, with celebrity footage that over-represents frontal poses and good illumination versus operational settings.Its degradation is synthetic, plate strings are Latin alphanumeric, and recognisers are used off the shelf.Source: §8 · p. 9

Ethics

• The study discourages operational surveillance or law-enforcement use without ethical review and compliance with local privacy law.Source: §8 · p. 9

Conclusion

• FaceRecBox rises from 0.686 to 0.722 and TextRecBox from 0.309 to 0.367 with the 2.3 M-parameter baseline; RCDM-RMGF further improves plate TextRecBox to 0.3801.The paper releases annotations, metadata, download and degradation scripts, and scorers for comparing VSR by recognisability.Source: §9 · p. 10

Data and code availability

• Annotations, gallery metadata, download scripts, degradation generators, two scorers, and the fidelity/correlation script are provided through the FANVID repository and accompanying code directory.Source: Data and code availability · p. 10

• Generative AI tools assisted with language editing, consistency checks, and manuscript preparation.Source: Declaration of generative AI and… · p. 10
• The authors reviewed and edited the manuscript and are responsible for its final content.Source: Declaration of generative AI and… · p. 10