Corruption-aware Feature Completion (CFC) Overview

• CFC is a feature-space paradigm that explicitly models corruption using masks, latent embeddings, and expert-guided residual refinement.
• It improves recovery by separating informative residuals from artifact-dominated features in tasks like video recovery and point cloud completion.
• Empirical results across domains show that sophisticated supervision and corruption-aware optimization lead to significant performance gains.

Searching arXiv for the named CFC work and closely related feature-completion/corruption-aware papers to ground the article. arXiv search query: all:"Corruption-aware Feature Completion" OR ti:"Corruption-aware Feature Completion" arXiv search query: ti:"Towards Blind Bitstream-corrupted Video Recovery via a Visual Foundation Model-driven Framework" Corruption-aware Feature Completion (CFC) denotes a class of completion and recovery methods that treats corruption as a first-class modeling object rather than as a nuisance to be ignored. In the formulation that introduces the term explicitly, CFC is the intermediate feature-processing module in blind bitstream-corrupted video recovery: it uses corruption masks, multi-scale foundational embeddings, and a CLIP-guided Mixture-of-Residual-Experts to enhance informative residuals while suppressing artifact-dominated ones (Liu et al., 30 Jul 2025). Closely related work applies the same principle to corrupted point clouds, occluded person re-identification, temporally corrupted skeleton sequences, low-rank matrix completion with sparse gross corruptions, and training under stochastic feature corruption, suggesting that CFC is best understood as a broader feature-space paradigm for corruption-aware inference rather than as a single architecture (Tesema et al., 22 Jul 2025).

1. Definition, scope, and conceptual lineage

CFC is characterized by three commitments. First, corruption is modeled explicitly, either through masks, sparse corruption variables, noise-specific feature branches, or analytically specified corruption distributions. Second, completion is performed in feature space or latent structure space, rather than only at the pixel or coordinate level. Third, the completed representation is constrained to remain useful for the downstream task, such as reconstruction, recovery, recognition, retrieval, or classification.

In blind bitstream-corrupted video recovery, the corruption is induced by packet loss, bit flips, or syntax errors in compressed bitstreams, producing large and irregular pixel-domain artifacts. The corresponding CFC module receives corruption features, DAC masks, multi-scale corruption embeddings from a Visual Foundation Model, a preliminary completed feature from a pre-existing BSCVR feature completion module, and a CLIP-based high-level corruption token. It then produces a refined corruption-aware feature that suppresses artifact-dominated channels and enhances informative residual channels (Liu et al., 30 Jul 2025). In this setting, “corruption-aware” means that feature refinement is conditioned on both where corruption is localized and what kind of corruption is semantically present.

The same structural idea appears in other domains under different names. DWCNet for point cloud completion inserts a Noise Management Module between encoder and decoder, explicitly splitting encoded features into clean and noisy components and feeding only the clean features to the completion decoder (Tesema et al., 22 Jul 2025). FCFormer and RFCnet for occluded person re-identification do not use the name CFC, but both reconstruct occluded semantics in feature space rather than discarding occluded regions, one via transformer completion tokens and one via spatial-temporal region completion (Wang et al., 2023); (Hou et al., 2021). FineTec similarly reconstructs a base skeleton sequence from temporally corrupted input by a mask-aware, context-aware completion module before recognition (Shao et al., 31 Dec 2025). Earlier antecedents are even more abstract: PARSuMi explicitly separates low-rank structure from sparse gross corruption in a partially observed matrix, while marginalized corrupted features trains predictors under the expectation over corrupted inputs rather than over only clean samples (Wang et al., 2013); (Maaten et al., 2014).

This suggests a broad technical definition: CFC is any method that completes or recovers task-relevant structure from partial or corrupted observations by explicitly separating reliable structure from corruption, and by using only the reliable or corrected component for downstream inference.

2. Problem formulations and corruption models

The formal object of completion varies by modality, but the recurring distinction is between latent structure and corruption process.

In bitstream-corrupted video recovery, the corrupted sequence is written as

$$X = \{x_i \in \mathbb{R}^{H \times W \times 3}\}_{i=1}^{L},$$

with associated bitstream $X_b$, and the objective is to recover

$\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$

that matches the clean sequence spatially and temporally. The corruptions are not simple degradations such as uniform blur or additive noise; they include wrong references, frozen regions, color stripes, missing macroblocks, and motion-compensated failures, and may retain partial residual information inside corrupted regions (Liu et al., 30 Jul 2025). This is precisely why corruption-aware residual modulation becomes necessary.

In robust point cloud completion, the input is a partial point cloud

$P \in \mathbb{R}^{N \times 3},$

and the task is to predict a dense complete point cloud of the underlying object. The corruption space is unusually rich: external object interference $E_{OI}$, background interference from wall $BI_W$ and floor $BI_F$, occlusion by other objects $O_{BOO}$, dynamic jitter with trajectory $D_{JT}$, triaxial rotation $T_R$, isometric scaling $X_b$0, and a randomly combined corruption $X_b$1 formed by applying a random subset of 2–7 corruption types jointly (Tesema et al., 22 Jul 2025). The corresponding CPCCD benchmark is designed precisely to expose the brittleness of models trained only on clean synthetic partials.

In practical matrix completion, the observed matrix $X_b$2 is only partially observed on an index set $X_b$3, while some observed entries are grossly corrupted and others are perturbed by dense small noise. PARSuMi formalizes this through an explicit low-rank matrix $X_b$4 and sparse corruption matrix $X_b$5:

$X_b$6

subject to $X_b$7, $X_b$8, $X_b$9, and $\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$0 (Wang et al., 2013). Here the completion problem is already corruption-aware by construction: missing entries, dense noise, and sparse gross corruptions are modeled as different phenomena.

In temporally corrupted skeleton recognition, the clean sequence is

$\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$1

while the corrupted sequence

$\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$2

contains only $\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$3 valid frames and uses zero-padding for missing frames. Corruption is simulated by random frame dropping at 25%, 50%, and 75%, corresponding to minor, moderate, and severe settings (Shao et al., 31 Dec 2025). The completion problem is therefore one of restoring temporal continuity and preserving fine-grained motion cues rather than removing additive noise.

In marginalized corrupted features, the input itself is treated as a random variable under a corruption model $\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$4, and learning minimizes

$\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$5

Corruptions include blankout, Gaussian noise, Laplace noise, and Poisson corruption (Maaten et al., 2014). Unlike explicit feature reconstruction, this formulation “completes” information at the decision level by training the predictor to remain correct under corrupted feature realizations.

3. Architectural patterns of corruption awareness

CFC systems differ sharply in modality, but several architectural motifs recur.

The first is explicit corruption localization or partitioning. In blind video recovery, Detect Any Corruption predicts corruption localization masks $\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$6 and multi-scale foundational embeddings $\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$7, which become the two principal side signals for CFC (Liu et al., 30 Jul 2025). In RFCnet, a foreground probability map distinguishes likely person regions from background or occluder regions, while an Adaptive Partition Unit divides the feature map into six semantically structured body regions (Hou et al., 2021). In FineTec, the mask indicator is embedded directly into the per-frame input representation, making temporal absence explicit (Shao et al., 31 Dec 2025).

The second is feature-space separation of structure and corruption. DWCNet’s Noise Management Module is a direct instance of this idea. The encoder output $\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$8 is sent through a clean path based on multi-head self-attention and feed-forward refinement, and a noisy path based on multi-scale 1D convolutions over the sequence dimension. The clean feature is pulled toward encoder features of clean partial point clouds, and pushed away from the noisy feature via contrastive loss. The decoder then consumes only $\hat{Y} = \{\hat{y}_i\}_{i=1}^{L}$9 (Tesema et al., 22 Jul 2025). This turns completion into reconstruction from a corruption-filtered latent subspace.

The third is contextual completion through latent tokens, clusters, or memory. FCFormer’s Feature Completion Decoder prepends learnable completion tokens to occluded feature tokens, performs self-attention, and produces completed features that approximate holistic features (Wang et al., 2023). RFCnet’s Spatial RFC clusters region features using appearance and position assignment matrices and reconstructs each region from those clusters, while Temporal RFC uses cross-frame attention and a gate to decide whether a region feature should be preserved or replaced by temporal context (Hou et al., 2021). FineTec’s completion module relies on in-context learning: a clean context sequence and a synthetically masked version serve as a prompt pair, while the actual corrupted input and a prior sequence form the query pair (Shao et al., 31 Dec 2025).

The fourth is high-level semantic control over residual reuse. In the named CFC module for video recovery, hierarchical augmentation aligns corruption features with VFM embeddings by scale-wise cross-attention, while MoRE uses a CLIP-based corruption embedding to gate multiple residual experts. A final corruption-aware residual enhancement step performs channel-wise recalibration through cross-attention between the CLIP token and refined residual features (Liu et al., 30 Jul 2025). This explicitly rejects the corruption-agnostic assumption that all residuals inside corrupted regions are equally useful.

The fifth is probabilistic or algebraic corruption modeling instead of neural routing. PARSuMi alternates between a rank-constrained low-rank update for $P \in \mathbb{R}^{N \times 3},$0 and an exact $P \in \mathbb{R}^{N \times 3},$1-constrained corruption update for $P \in \mathbb{R}^{N \times 3},$2, so corruption awareness is enforced by optimization structure rather than by learned masks or attention (Wang et al., 2013). MCF instead integrates over the corruption distribution analytically, transforming robustness into an expected-risk objective (Maaten et al., 2014). These are not feature-completion decoders in the transformer sense, but they instantiate the same separation between clean structure and corruption process.

4. Supervision, objectives, and optimization

CFC methods are defined as much by their supervision scheme as by their architecture.

In DWCNet, the total loss is

$P \in \mathbb{R}^{N \times 3},$3

where $P \in \mathbb{R}^{N \times 3},$4 is the Chamfer Distance between predicted and ground-truth point clouds, and $P \in \mathbb{R}^{N \times 3},$5 combines a positive term aligning $P \in \mathbb{R}^{N \times 3},$6 with clean-partial features and a negative term penalizing alignment between $P \in \mathbb{R}^{N \times 3},$7 and $P \in \mathbb{R}^{N \times 3},$8 (Tesema et al., 22 Jul 2025). This is explicit disentanglement supervision.

FCFormer uses a more classically discriminative package:

$P \in \mathbb{R}^{N \times 3},$9

Here $E_{OI}$0 is a direct feature-space regression from occluded to holistic tokens, $E_{OI}$1 is a cross-hard triplet loss over holistic, occluded, and completed features, and $E_{OI}$2 aligns the identity distributions of completed and holistic features (Wang et al., 2023). Completion is therefore constrained geometrically, metrically, and distributionally.

FineTec pre-trains its completion module with

$E_{OI}$3

using both query reconstruction and prompt reconstruction. Downstream recognition then adds

$E_{OI}$4

where $E_{OI}$5 is an acceleration reconstruction loss derived from physics-driven estimation (Shao et al., 31 Dec 2025). The result is a two-stage arrangement: completion is first trained to restore plausible motion, and recognition is later trained on the completed sequence.

The explicit CFC module in bitstream-corrupted video recovery is optimized only through the reconstruction objectives of the frozen BSCVR recovery network. The paper states that there is no explicit supervision on gating weights or expert specialization; expert routing, hierarchical augmentation, and channel recalibration are learned through their effect on frame-wise, perceptual, and temporal reconstruction quality (Liu et al., 30 Jul 2025). This makes CFC an implicit corruption-aware optimizer embedded inside a larger recovery system.

PARSuMi employs proximal alternating minimization. The $E_{OI}$6-step enforces $E_{OI}$7 and is solved via a subspace parameterization with LM_GN plus a quadratic-majorization safeguard; the $E_{OI}$8-step has a closed-form exact solution under $E_{OI}$9 and Frobenius constraints by selecting the top-$BI_W$0 residual entries (Wang et al., 2013). MCF, by contrast, optimizes an analytically marginalized objective under a specified corruption model, giving closed-form solutions for quadratic loss and convex objectives or bounds for exponential and logistic loss (Maaten et al., 2014). Taken together, these works show that CFC can be supervised by direct reconstruction, clean–noisy contrast, distribution matching, metric learning, constrained alternating optimization, or marginalized robust risk.

5. Representative systems and empirical evidence

Across domains, empirical results consistently support the claim that explicit corruption awareness is not interchangeable with ordinary completion.

Domain	Representative system	Reported effect
Bitstream-corrupted video	DAC + CFC + MoRE	Improves over BSCVR-P* in oracle-mask setting
Point cloud completion	DWCNet + NMM	Best or near-best CD-L1 across clean and corrupted CPCCD categories
Occluded person Re-ID	FCFormer	Improves Rank-1 and mAP on occluded benchmarks
Occluded video/image Re-ID	RFCnet	Large gains over baselines on Occluded-Duke settings
Skeleton action recognition	FineTec	Better restoration and higher Top-1 under severe temporal corruption
Matrix completion / robust learning	PARSuMi, MCF	Explicitly separates gross corruption from structure

In blind bitstream-corrupted video recovery, the non-blind oracle-mask comparison isolates the effect of CFC itself. On the YouTube-VOS subset, BSCVR-P* reports PSNR 31.56, SSIM 0.9536, LPIPS 0.0288, and VFID 0.0296, whereas Ours* reports PSNR 32.03, SSIM 0.9605, LPIPS 0.0279, and VFID 0.0286. On the DAVIS subset, BSCVR-P* gives PSNR 28.29, SSIM 0.9147, LPIPS 0.0395, and VFID 0.1484, while Ours* gives PSNR 28.60, SSIM 0.9207, LPIPS 0.0403, and VFID 0.1461. The paper’s ablations further indicate that MoRE provides the major gain, while hierarchical augmentation yields a smaller additional PSNR/SSIM improvement (Liu et al., 30 Jul 2025).

In corrupted point cloud completion, clean-only training produces severe degradation on CPCCD. After fine-tuning on the $BI_W$1 subset, DWCNet reaches CD-L1 9.754 on the hardest corruption $BI_W$2, compared with 10.317 for AdaPoinTr, and 7.727 on clean PCN compared with 8.150 for AdaPoinTr. The NMM ablation is especially diagnostic: on $BI_W$3, DWCNet without NMM gives CD-L1 13.403, whereas DWCNet with NMM gives 10.263 after 200 epochs (Tesema et al., 22 Jul 2025). The paper’s path ablations also show that the attention-based clean path is the dominant contributor and that replacing it with an MLP causes a large drop.

For occluded person Re-ID, FCFormer reports Rank-1 71.3% and mAP 60.9% on Occluded-Duke, with 73.0%/63.1% under the small sliding-window stride variant, and 79.4%/77.2% with re-ranking. On P-DukeMTMC it reports Rank-1 91.5% and mAP 80.7%, and on Occluded-REID it reports Rank-1 84.9% and mAP 86.2% (Wang et al., 2023). Its ablations show additive benefits from OIA, the dual stream, FCD, CHT, and FC$BI_W$4, consistent with the interpretation that realistic corruption modeling and feature completion are complementary rather than redundant.

RFCnet provides earlier evidence for feature-space completion under occlusion. On Occluded-DukeMTMC, the full model reports mAP 54.5 and Rank-1 63.9. On Occluded-DukeMTMC-VideoReID, RFCnet with pose and foreground supervision reports mAP 92.0 and Rank-1 93.0. The ablations show that SRFC and TRFC each outperform the baseline individually, and that their sequential combination performs best (Hou et al., 2021).

In temporally corrupted skeleton recognition, FineTec reports top-1 accuracies of 89.1% on Gym99-severe and 78.1% on Gym288-severe, and its completion module reduces severe-corruption MPJPE from 0.192 for the best baseline to 0.147. The ablation without in-context learning yields severe MPJPE 0.169 instead of 0.147, directly supporting the claim that context-aware completion is central rather than incidental (Shao et al., 31 Dec 2025).

Earlier robust formulations support the same thesis in different language. PARSuMi is reported to operate successfully in a much larger range of practical problems than convex alternatives and to detect hidden corruptions in SfM and photometric stereo settings (Wang et al., 2013). MCF shows strong robustness under feature deletion at test time and interprets robustness as training under the expectation of corrupted inputs, rather than by explicit imputation (Maaten et al., 2014).

6. Relation to adjacent paradigms, misconceptions, and open directions

CFC is often conflated with mask-based inpainting, generic denoising, or data augmentation, but the literature draws sharper distinctions. In blind bitstream-corrupted video recovery, classical inpainting and error concealment typically discard residuals inside corrupted regions and treat the region as unknown, whereas CFC selectively exploits residual information and suppresses artifact-induced residuals (Liu et al., 30 Jul 2025). In point cloud completion, the central claim is that robustness comes not only from data augmentation, but from explicitly modeling corruption at the feature level and supervising that separation (Tesema et al., 22 Jul 2025). In occluded Re-ID, both FCFormer and RFCnet reject the common strategy of simply ignoring occluded regions and instead attempt latent semantic recovery (Wang et al., 2023); (Hou et al., 2021).

A second misconception is that CFC requires explicit clean targets at every stage. Some methods do use direct clean supervision, such as FCFormer’s feature regression to holistic features or DWCNet’s alignment to clean partial features. Others do not: the video CFC module learns expert routing without explicit gate supervision, and MCF never reconstructs features at all, instead marginalizing the corruption distribution in the predictive objective (Liu et al., 30 Jul 2025); (Maaten et al., 2014). This suggests that CFC is compatible with both explicit reconstruction and implicit robust inference.

The main limitations are domain-dependent. The VFM-driven video framework is computationally heavy and depends on SAM2.1, DINOv2, and CLIP embeddings; rare or out-of-distribution corruptions may still cause DAC mis-localization (Liu et al., 30 Jul 2025). DWCNet’s CPCCD mimics indoor real scans but does not capture all sensor physics and remains limited to PCN’s eight categories (Tesema et al., 22 Jul 2025). PARSuMi has scalability and initialization sensitivity issues, while MCF assumes factorized corruption models and remains focused on linear predictors (Wang et al., 2013); (Maaten et al., 2014). Re-ID formulations are specialized to the structured human body and to occlusion as corruption (Wang et al., 2023); (Hou et al., 2021). FineTec’s success depends on a large skeleton bank and on the adequacy of the simulated masking patterns (Shao et al., 31 Dec 2025).

Open directions stated in the literature converge on a common agenda: lighter or distilled foundation models for corruption-aware recovery; richer corruption simulators including sensor physics, adversarial corruptions, temporal misalignment, and multimodal degradation; more explicit expert-specialization or latent disentangling losses; end-to-end co-training of corruption localization and feature completion; broader benchmarks spanning scenes, outdoor LiDAR, or multimodal restoration; and extension of the CFC principle to other local, irregular degradations (Liu et al., 30 Jul 2025); (Tesema et al., 22 Jul 2025). A plausible implication is that future CFC systems will be judged less by whether they “fill in” missing data visually and more by whether they maintain task-consistent latent structure under severe distribution shift.