ProCap: Diverse Methods in Vision & Compression

• ProCap is a versatile term applied in multiple domains, including change captioning with explicit procedure modeling, spatial augmented reality with dual captioning, and CNN compression via cascaded projection.
• In change captioning, ProCap utilizes a two-stage framework that models dynamic transformations through learned procedure queries to overcome static comparison limitations.
• For CNN compression, ProCap employs cascaded projection with proxy matrix optimization, reducing computation and memory while preserving classification accuracy.

ProCap is a name used for multiple unrelated research systems in recent arXiv literature. In change captioning, "Imagine How To Change: Explicit Procedure Modeling for Change Captioning" defines ProCap as a two-stage framework that shifts image change understanding from static comparison to dynamic procedure modeling (Sun et al., 6 Mar 2026). In spatial augmented reality, "ProCap: Projection-Aware Captioning for Spatial Augmented Reality" defines ProCap as a projection-aware dual-captioning pipeline that separates physical scene semantics from projected content (Cao et al., 1 Apr 2026). In CNN compression, Cascaded Projection (CaP), also referred to as ProCap in some contexts, denotes an end-to-end method for compressing successive convolutional layers through a shared low-dimensional projection (Minnehan et al., 2019). A related communications paper states that it does not mention “ProCap” explicitly, but that one can interpret “ProCap” as a programmable, processing-capable continuous aperture in CAPA-based 6G architectures (Liu et al., 2024).

1. Terminological scope

In current arXiv usage, ProCap is not a single method but a recurring name attached to distinct technical programs. The term spans multimodal change understanding, spatial augmented reality, and projection-based model compression, while the CAPA literature uses it only as an interpretive extension rather than as a formal method name.

Usage of ProCap	Problem domain	Core mechanism
ProCap	Change captioning	Explicit procedure modeling, procedure encoder, learnable procedure queries
ProCap	Spatial augmented reality	Automated segmentation, region-aware retrieval, dual captioning
CaP / ProCap	CNN compression	Cascaded projection, proxy matrix optimization, kernel relaxation
ProCap as interpretive term	CAPA-based 6G systems	Programmable, processing-capable continuous aperture

The three established usages are technically distinct. The change-captioning system models a latent trajectory between $I_\text{bef}$ and $I_\text{aft}$. The SAR system decomposes a composite image into physical-scene and projection layers. The compression method projects the output channels of layer $i$ and the input channels of layer $i+1$ into a unified low-dimensional space. The CAPA discussion is adjacent rather than canonical: it proposes no paper titled ProCap, but explicitly frames ProCap as a useful interpretation of programmable continuous apertures.

2. ProCap in change captioning

In change captioning, the input is a pair of images of the same scene taken at different times,

$(I_\text{bef}, I_\text{aft}),$

and the goal is to generate a natural language description $T$ focusing on what changed. ProCap reformulates this from static image comparison into dynamic procedure modeling, where a valid change procedure with respect to a caption $T$ is a continuous mapping

$\gamma_T: [0,1] \to \mathcal{I},$

with $\gamma_T(0)=I_\text{bef}$, $\gamma_T(1)=I_\text{aft}$, continuous evolution of object attributes $I_\text{aft}$0, consistency with the semantic constraints in $I_\text{aft}$1, and invariance of unchanged objects over time (Sun et al., 6 Mar 2026).

The motivation is that static image pairs obscure temporal ambiguity, viewpoint and distractor effects, and multi-step or composite changes. Prior non-LLM methods such as DUDA, NCT, VARD-Trans, and SMART concentrated on difference extractors and alignment modules, while LLM-based methods such as Qwen-VL, LLaVA, and FINER retained essentially the same static-pair visual interface. ProCap addresses this with a two-stage framework. Stage 1, Explicit Procedure Modeling (EPM), generates a dense sequence of interpolated frames using VFIformer, scores them for informativeness, samples a sparse set of keyframes, and trains a Transformer-based procedure encoder with caption-conditioned masked sequence modeling, cross-modal alignment, and temporal consistency losses. Stage 2, Implicit Procedure Captioning (IPC), reuses the trained encoder in an encoder-decoder captioning model and replaces explicit intermediate frames at inference with $I_\text{aft}$2 learnable procedure query tokens.

The discrete approximation to the procedure is

$I_\text{aft}$3

Dense pseudo-procedures are generated with recursive frame interpolation, with $I_\text{aft}$4 in practice. Because recursive interpolation introduces temporal redundancy, ProCap uses a confidence score

$I_\text{aft}$5

and selects the top-$I_\text{aft}$6 frames, typically $I_\text{aft}$7. The similarity function is instantiated either as visual-only similarity with DINOv2 ViT-L/14 or as visual-text similarity with CLIP4IDC; the visual+text variant is reported as more robust because it penalizes visually plausible but caption-inconsistent frames.

The procedure encoder uses a frozen VQGAN encoder. Images are resized to $I_\text{aft}$8, encoded to a $I_\text{aft}$9 latent grid, and yield $i$0 patch features per frame. The embedding dimension is $i$1. The full encoder input is

$i$2

where $i$3 is a learnable token for cross-modal alignment and $i$4 is a learnable token for temporal consistency. Only visual tokens are masked. The masking schemes are Entire masking, Random patch masking, In-block masking, and Out-of-block masking, sampled with probabilities $i$5. The procedure modeling loss is

$i$6

Here $i$7 is VQ-VAE-style discrete reconstruction conditioned on language, $i$8 distinguishes matched from mismatched captions, and $i$9 distinguishes temporally coherent from incoherent sequences. Stage 2 retains the encoder, inserts $i+1$0 learnable procedure queries initialized as the mask embedding $i+1$1, and trains a Transformer-based autoregressive text decoder with hidden size $i+1$2 using

$i+1$3

The framework is evaluated on CLEVR-Change, Spot-the-Diff, and Image-Editing-Request. CLEVR-Change contains 79,606 image pairs and 493,735 change captions; Spot-the-Diff contains 13,192 image pairs with human captions; Image-Editing-Request contains 3,939 pairs and 5,695 captions. On CLEVR-Change, ProCap reports $i+1$4, $i+1$5, $i+1$6, $i+1$7. On Spot-the-Diff it reports $i+1$8, $i+1$9, $(I_\text{bef}, I_\text{aft}),$0, $(I_\text{bef}, I_\text{aft}),$1. On Image-Editing-Request it reports $(I_\text{bef}, I_\text{aft}),$2, $(I_\text{bef}, I_\text{aft}),$3, $(I_\text{bef}, I_\text{aft}),$4, $(I_\text{bef}, I_\text{aft}),$5. Category-wise on CLEVR-Change, ProCap is best on Add, Drop, and Move, with Move at $(I_\text{bef}, I_\text{aft}),$6 METEOR versus $(I_\text{bef}, I_\text{aft}),$7 for DIRL+CCR. Ablations attribute much of the gain to the combination of Stage 1 pretraining and procedure queries: on CLEVR-Change, a baseline static encoder-decoder gives $(I_\text{bef}, I_\text{aft}),$8, EPM only gives $(I_\text{bef}, I_\text{aft}),$9, queries only with $T$0 give $T$1, and full ProCap with $T$2 gives $T$3. With $T$4, CIDEr peaks at $T$5 and Tokens Per Second is approximately $T$6, while explicit procedure captioning at inference is both slower and worse, with CIDEr $T$7 and TPS $T$8. The authors also report code and pre-trained models at the project repository.

3. ProCap in spatial augmented reality

In spatial augmented reality, a camera observes a composite image in which some pixels come from the physical scene and others show projected digital content distorted by geometry, surface reflectance, and lighting. ProCap treats SAR understanding as projection-aware dual captioning. The observed image is modeled as

$T$9

where $T$0 is the physical scene appearance, $T$1 is the projected content, and $T$2 is a complex blending operator. The task is

$T$3

where $T$4 describes the physical scene excluding projected content and $T$5 describes the projected content excluding ambient context (Cao et al., 1 Apr 2026).

The paper identifies three failure modes of standard VLMs on SAR imagery: virtual-physical ambiguity, perceptual degradation, and lack of SAR-specific semantic data and metrics. To address them, ProCap employs a two-stage pipeline. Stage 1 uses a frozen CLIP ViT-g encoder (EVA-CLIP) to extract coarse features $T$6, two deconvolution layers $T$7 for refinement, and a lightweight two-layer convolutional segmentation module $T$8 to predict a binary projection mask

$T$9

Ground-truth masks are derived per scene using a uniform white projection and are intentionally coarse. Stage 2 performs mask pooling, region-aware retrieval, and dual captioning. Projection-specific features are

$\gamma_T: [0,1] \to \mathcal{I},$0

Two separate BLIP-2-style Q-Formers $\gamma_T: [0,1] \to \mathcal{I},$1 and $\gamma_T: [0,1] \to \mathcal{I},$2 produce scene and projection embeddings $\gamma_T: [0,1] \to \mathcal{I},$3 and $\gamma_T: [0,1] \to \mathcal{I},$4. A knowledge base

$\gamma_T: [0,1] \to \mathcal{I},$5

stores LVIS visual embeddings and textual labels. Retrieval computes cosine similarity between $\gamma_T: [0,1] \to \mathcal{I},$6 and $\gamma_T: [0,1] \to \mathcal{I},$7, retrieves top-$\gamma_T: [0,1] \to \mathcal{I},$8 names with $\gamma_T: [0,1] \to \mathcal{I},$9, and fuses them with the projection embedding via a knowledge Q-Former $\gamma_T(0)=I_\text{bef}$0. A linear layer $\gamma_T(0)=I_\text{bef}$1 maps features into the input space of a frozen LLM decoder, and task-specific conditioning is formed as

$\gamma_T(0)=I_\text{bef}$2

Training uses dual caption losses and segmentation BCE: $\gamma_T(0)=I_\text{bef}$3

$\gamma_T(0)=I_\text{bef}$4

$\gamma_T(0)=I_\text{bef}$5

and

$\gamma_T(0)=I_\text{bef}$6

with $\gamma_T(0)=I_\text{bef}$7, $\gamma_T(0)=I_\text{bef}$8, and $\gamma_T(0)=I_\text{bef}$9. The LLM backbones in the reported variants include TinyLlama-1.1B, OPT-2.7B, OpenLLaMA-3B, Vicuna-7B, and Llama-3.3-8B, with the LLM parameters frozen. An additional experiment fine-tunes Qwen3-VL-8B-Instruct on RGBP.

RGBP is introduced as the first large-scale SAR semantic benchmark dataset. It contains 65 diverse physical scenes and 180,678 SAR image pairs. Projection content is drawn from COCO 2017, nocaps, and WHOOPS!. The train split uses 60 physical scenes and 118,287 pairs total. The evaluation split covers 65 scenes, with 60 seen scenes and 5 unseen scenes, and contains 62,400 image pairs. Ambient illumination ranges from approximately 2.9 to 546 lux. Each SAR image is annotated with a coarse binary projection mask and two captions: one for the scene and one for the projection. Evaluation uses explicit natural-language prompts together with internal task-specific tokens $\gamma_T(1)=I_\text{aft}$0 and $\gamma_T(1)=I_\text{aft}$1, and reports BLEU@4, METEOR, CIDEr, and SPICE separately for scene and projection captioning.

The main findings are strongly asymmetric between scene and projection captioning. On seen scenes, off-the-shelf VLMs have scene-captioning CIDEr around $\gamma_T(1)=I_\text{aft}$2 to $\gamma_T(1)=I_\text{aft}$3, while ProCap$\gamma_T(1)=I_\text{aft}$4 reaches CIDEr $\gamma_T(1)=I_\text{aft}$5 on COCO and $\gamma_T(1)=I_\text{aft}$6 on WHOOPS!. For projection captioning on seen scenes, off-the-shelf VLMs have CIDEr around $\gamma_T(1)=I_\text{aft}$7 to $\gamma_T(1)=I_\text{aft}$8 on COCO, whereas ProCap$\gamma_T(1)=I_\text{aft}$9 reaches CIDEr $I_\text{aft}$00 on COCO and ProCap$I_\text{aft}$01 reaches $I_\text{aft}$02. Fine-tuned Qwen3-VL-8B-Instruct$I_\text{aft}$03 reaches projection CIDEr $I_\text{aft}$04 on COCO and $I_\text{aft}$05 on WHOOPS!, with SPICE $I_\text{aft}$06 on COCO. On unseen scenes, ProCap$I_\text{aft}$07 reports projection CIDEr $I_\text{aft}$08 on COCO, ProCap$I_\text{aft}$09 reports $I_\text{aft}$10, and Qwen3-VL-8B-Instruct$I_\text{aft}$11 reports $I_\text{aft}$12. Ablations isolate the effect of each module. Removing feature refinement reduces projection CIDEr from $I_\text{aft}$13 to $I_\text{aft}$14 on unseen-scene COCO. Removing region-aware retrieval reduces projection CIDEr from $I_\text{aft}$15 to $I_\text{aft}$16. Mask-pooling ablations show a trade-off between strict decoupling and some caption metrics, but the authors retain masking to preserve reliable dual evaluation. Source code, pre-trained models, and RGBP are released on the project page.

4. CaP / ProCap in convolutional network compression

Cascaded Projection (CaP), also referred to as ProCap in some contexts, addresses compression and acceleration of deep convolutional networks such as VGG16 and ResNets. The objective is to reduce FLOPs, reduce parameter count and memory, and preserve classification accuracy as much as possible, while working end-to-end with SGD and backpropagation (Minnehan et al., 2019).

The method is positioned between low-rank feature factorization and channel pruning. Low-rank feature methods reduce computation but require reprojection back to the original dimensionality, which increases memory traffic. Channel pruning avoids reprojection but removes entire filters and channels, which can discard substantial information. CaP instead forms linear combinations of channels and pulls the next layer’s kernels down into the same low-dimensional space. For convolutional layer $I_\text{aft}$17, with input tensor $I_\text{aft}$18, weights $I_\text{aft}$19, and bias $I_\text{aft}$20, CaP introduces a projection matrix

$I_\text{aft}$21

with orthogonality constraint

$I_\text{aft}$22

The compressed current-layer filters are

$I_\text{aft}$23

and the next-layer filters become

$I_\text{aft}$24

Because $I_\text{aft}$25 is absorbed into $I_\text{aft}$26, there is no explicit reprojection layer in the forward pass. The output and the following layer’s input thus live in a unified low-dimensional space.

For two successive layers, the reconstruction loss compares the original output after two layers to the compressed output: $I_\text{aft}$27 This is combined with classification loss as

$I_\text{aft}$28

For multi-layer compression, deep supervision sums reconstruction losses across layers. The orthogonality constraint makes direct optimization non-convex, so CaP introduces a proxy matrix $I_\text{aft}$29. If

$I_\text{aft}$30

then the projection is defined by

$I_\text{aft}$31

This Proxy Matrix Projection mechanism keeps $I_\text{aft}$32 on the Grassmann manifold while allowing standard backpropagation. After projection optimization, the compressed kernels are materialized and the explicit projection layers are removed. A subsequent kernel relaxation stage fine-tunes only the compressed kernels.

The method reports strong results on CIFAR and ImageNet. On CIFAR-100 with ResNet18 at 50% reduction, the uncompressed baseline is $I_\text{aft}$33, a compressed ResNet18 trained from scratch gives $I_\text{aft}$34, CaP with projection-only gives $I_\text{aft}$35, random projections with kernel relaxation give $I_\text{aft}$36, and optimized projections with kernel relaxation give $I_\text{aft}$37. On CIFAR-10, ResNet56 with approximately 50.2% FLOPs reports $I_\text{aft}$38 versus $I_\text{aft}$39 baseline without fine-tuning, and $I_\text{aft}$40 versus $I_\text{aft}$41 with fine-tuning. ResNet110 with approximately 50.1% FLOPs reports $I_\text{aft}$42 versus $I_\text{aft}$43 baseline without fine-tuning, and $I_\text{aft}$44 versus $I_\text{aft}$45 with fine-tuning. For VGG16 on ImageNet, the headline result is approximately $I_\text{aft}$46 FLOP reduction. CaP Optimal, without fine-tuning, reports FLOPs $I_\text{aft}$47B, parameters $I_\text{aft}$48M, memory $I_\text{aft}$49 MB, GPU speedup $I_\text{aft}$50, and Top-5 $I_\text{aft}$51 versus $I_\text{aft}$52 baseline. With fine-tuning, CaP reports FLOPs $I_\text{aft}$53B, memory $I_\text{aft}$54 MB, and Top-5 $I_\text{aft}$55 versus $I_\text{aft}$56 baseline. The paper also notes limitations: the choice of ranks $I_\text{aft}$57 is not automatic, certain critical structures such as the outputs of the last convolution in each residual block and the skip connections are not compressed, and training with SVD-based projection introduces overhead.

5. Shared structural themes and major differences

The data suggest that the repeated reuse of the name ProCap reflects a common design instinct rather than a shared domain. In the change-captioning system, the ambiguous object is the latent temporal trajectory between two still images. In the SAR system, the ambiguous object is the mixture of physical scene and projected content in a single RGB frame. In CaP, the ambiguous object is the redundant intermediate channel space between consecutive convolutional layers. Each method inserts an intermediate structure that is easier to constrain: keyframe procedures and learnable procedure queries in change captioning, coarse masks plus retrieved semantic names in SAR, and orthogonal low-dimensional projections in CNN compression (Sun et al., 6 Mar 2026, Cao et al., 1 Apr 2026, Minnehan et al., 2019).

The supervision regimes are correspondingly different. Change-captioning ProCap trains a procedure encoder with $I_\text{aft}$58, $I_\text{aft}$59, and $I_\text{aft}$60, then fine-tunes the encoder-decoder end-to-end with $I_\text{aft}$61. SAR ProCap uses $I_\text{aft}$62, $I_\text{aft}$63, and $I_\text{aft}$64, while freezing both EVA-CLIP and the LLM and training the connecting adapters, Q-Formers, and segmentation components. CaP combines reconstruction and classification losses under orthogonality constraints and then removes the projection machinery from the deployed model. This suggests that “ProCap” has functioned less as a stable acronym family than as a label for methods that resolve a difficult problem by explicit structural decoupling.

Their inference profiles also diverge sharply. Change-captioning ProCap avoids runtime frame synthesis by replacing explicit pseudo-frames with learnable procedure queries, and the reported comparison shows implicit captioning is both faster and more accurate than explicit captioning. SAR ProCap retains a two-stage runtime consisting of segmentation, mask pooling, retrieval, and dual decoding because the decomposition itself is the task. CaP has the opposite objective: the projection apparatus exists during optimization, but after kernel compression and kernel relaxation the final inference graph is simply a standard CNN with fewer channels. The shared name therefore does not indicate interoperability or a common software stack.

6. Related programmable-aperture interpretation in CAPA literature

A separate line of work on continuous-aperture arrays uses ProCap only as an interpretive concept. The CAPA paper explicitly states that it does not mention “ProCap” explicitly, but that one can interpret “ProCap” as a programmable, processing-capable continuous aperture. In that interpretation, CAPA provides the mathematical backbone: the aperture current is modeled as a continuous function $I_\text{aft}$65, far-field radiation is given by a continuous-space Fourier transform,

$I_\text{aft}$66

and the CAPA channel is a compact integral operator with Hilbert-Schmidt decomposition (Liu et al., 2024).

The CAPA architecture is motivated by 6G trends toward electrically large apertures with continuous current distributions rather than spatially discrete arrays. Three hardware implementation approaches are identified: electrically driven metasurface-based leaky-wave antennas (MLWA), optically driven tightly coupled arrays (OTCA), and acoustically driven grating antennas (ITGA). Beamforming is formulated as functional optimization over continuous current distributions, with three solution families: discretization in the wavenumber domain, calculus of variations, and a subspace method in which the optimal beamformer lies in the span of user spatial responses,

$I_\text{aft}$67

The paper reports that CoV-based optimal and subspace-based designs are practically identical to optimal, near-optimal, and that all CAPA designs significantly outperform SPDAs, with the performance gap growing with aperture size.

In this interpretation, ProCap is not an established method name but a way to describe a programmable continuous aperture with embedded signal-processing capability. The CAPA paper uses this to map ideas such as local processing, analog spatial transforms, and general spatial operators onto a continuous-aperture framework. A plausible implication is that this is the loosest use of the term among the cited works: unlike the two 2026 vision papers and the CaP compression method, it functions as a conceptual bridge rather than a named architecture.

7. Status, limitations, and research significance

Across its established usages, ProCap names methods that were introduced to overcome a specific failure mode of more direct baselines. In change captioning, the failure mode is static-pair reasoning that cannot directly represent temporal dynamics or distributions over plausible procedures. In SAR, the failure mode is virtual-physical ambiguity and projection distortion that cause standard VLMs to hallucinate or merge layers. In CNN compression, the failure mode is the trade-off between low-rank factorization that increases memory traffic and channel pruning that discards entire filters (Sun et al., 6 Mar 2026, Cao et al., 1 Apr 2026, Minnehan et al., 2019).

The limitations are likewise domain-specific. Change-captioning ProCap depends on optical-flow-based frame interpolation, which is ill-posed under large viewpoint changes or occlusions, and the paper identifies subtle changes, noisy procedure reconstruction, and complex edits in open-ended IER as failure cases. SAR ProCap depends on coarse projection masks and on LVIS-based visual-name memory; the authors note degradation with inaccurate masks, limited knowledge-base coverage, modest scene diversity, and the use of earlier LLM backbones rather than newer systems such as Qwen3-VL. CaP depends on access to training data, manual or heuristic rank selection, and selective exclusion of sensitive residual structures from compression.

Taken together, the literature supports a precise but plural definition. ProCap is not a unified research program; it is an overloaded technical label attached to methods that make hidden structure explicit before optimization or caption generation. In the strongest sense, the term currently denotes two independent multimodal captioning frameworks and one projection-based network compression method. In a weaker, interpretive sense, it also names a way of thinking about programmable continuous apertures in 6G system design.