CATS Net: Multi-Domain Deep Learning Frameworks

• CATS Net is a suite of deep learning frameworks employing attention fusion, dual encoders, and graph-based reasoning to enhance diverse applications.
• It integrates specialized modules like tracing loss, CoFusion blocks, and temporal graph attention layers to deliver state-of-the-art performance in edge detection, medical segmentation, and security in V2X systems.
• The architectures showcase cross-domain impact by combining multi-scale feature fusion and consensus methods, achieving significant accuracy improvements and computational efficiency.

CATS Net refers to a collection of distinct architectures and frameworks—unrelated except for the shared acronym "CATS"—that have each introduced significant contributions in deep learning for edge detection, medical image segmentation, trust and security in cooperative autonomous systems, visual correspondence, temporal graph reasoning, time series forecasting, and domain-adaptive time series modeling. The following summary surveys the major CATS Net systems, delineating their mathematical foundations and empirical performance across several research domains.

1. Context-Aware Tracing Strategy for Crisp Edge Detection

The original CATS Net, introduced as the Context-Aware Tracing Strategy in edge detection, addresses the spatial localization ambiguity endemic to deep CNN-based edge detectors, particularly the issue of convolutional feature mixing. The architecture augments a multi-side-output CNN backbone (e.g., VGG16) with two modules: (1) a tracing loss for feature unmixing and (2) a pixelwise context-aware fusion (CoFusion) block (Huan et al., 2020).

Tracing Loss:

The full tracing loss is defined as:

$$L_{\rm trace}(Ŷ, Y) = L_{\rm ce}(Ŷ, Y) + \lambda_1 L_{\rm bdry}(Ŷ, Y) + \lambda_2 L_{\rm tex}(Ŷ, Y)$$

where $L_{\rm ce}$ is a weighted cross-entropy, $L_{\rm bdry}$ enforces mass concentration on annotated edge pixels over local patches, and $L_{\rm tex}$ penalizes spurious responses in texture zones.

CoFusion Block:

Multi-level side outputs $Z_1,\ldots,Z_L$ are fused using per-pixel self-attention:

$$w_{ij\ell} = \frac{\exp(a_{ij\ell})}{\sum_k \exp(a_{ijk})}, \qquad \mathrm{CoFusion}(Z)_{ij} = \sum_{\ell=1}^L w_{ij\ell}Z_{\ell,ij}$$

This enables dynamic weighting of crisp low-level versus robust high-level features.

Empirical Findings:

CATS yields significant "crispness" ODS improvements on BSDS500: RCF baseline 0.585 → 0.705 (+12%) without NMS, and similarly for BDCN (+6%). Ablations confirm complementarity of the boundary tracing and CoFusion modules (Huan et al., 2020).

2. CATS Architectures for Medical Image Segmentation

a. Complementary CNN and Transformer Encoders (CATS)

CATS for 3D medical segmentation leverages parallel CNN (3D U-Net) and Transformer encoders, fusing their representations by element-wise addition at each of four resolution levels before decoder up-sampling (Li et al., 2022). Multi-scale patch embeddings and skip-connected fusion are central to capturing both local spatial detail and global context.

Loss: Dice loss is:

$$\mathcal L_{\rm Dice} = 1 - \frac{2\sum_{i,k} p_{i,k}g_{i,k}}{\sum_{i,k} p_{i,k} + \sum_{i,k} g_{i,k}}$$

Performance:

CATS exceeds UNETR and CNN baselines on BTCV (81.4% vs 77.5–76.9% mean Dice), MoDA vestibular schwannoma (Dice 0.873; ASD 0.48 mm), and Decathlon-5 prostate (mean Dice 0.7877) (Li et al., 2022).

b. CATS v2: Hybrid CNN and Swin Transformer Encoders

CATS v2 extends this paradigm with a parallel Swin Transformer path employing shifted windowed multi-head self-attention. Fused features pass to the decoder via skip connections, and aggregation uses 1×1×1 convolutions for channel dimension parity (Li et al., 2023).

Empirical Highlights:

• BTCV: Dice 82.2% (CATS v2), outperforming Swin UNETR and original CATS.
• CrossMoDA: Dice 0.886 ± 0.076, ASD ≈ 0.48 mm.
• MSD-5: Dice 0.8034, best PZ and TZ scores (Li et al., 2023).

3. CATS for Secure Cooperative Autonomy in V2X Perception

The CATS framework for Cooperative Autonomy Trust & Security blends majority-based and reputation-based mechanisms for robust, privacy-preserving V2X sensor data sharing among autonomous vehicles (Asavisanu et al., 1 Mar 2025).

System Architecture:

• On-vehicle: Data broadcast with VPC-signed messages, in-situ verification, majority-view consensus, voting client for trust evidence.
• Central Security Authority (SA): Vote aggregation, reputation score update, trust state rollovers (trusted/untrusted/banned).
• Privacy: Rotating pseudonym certificates; no geolocation logging.

Key Mathematical Model:

Probability that $m$ trusted neighbors yield a wrong decision:

$$P_{WD}^{trusted}(m, P_{\mathrm{false}}) = \sum_{k=\lfloor m/2 + 1 \rfloor}^m \binom{m}{k} P_{\mathrm{false}}^k (1 - P_{\mathrm{false}})^{m-k}$$

Empirical Results:

• 230× reduction in false negatives vs. reputation-only baselines for message filtering.
• False positives <1%; mean ban time <58 s in city-scale simulations.
• Theoretical risk model yields $\overline{P_{WD}^{all}} < 10^{-6}$ under real sensor error rates (Asavisanu et al., 1 Mar 2025).

4. CATS in Temporal Graphs for Audio-Visual Video Parsing

The Category-Aware Temporal Graph (CATS) module enables segment-level semantic information propagation for weakly supervised audio-visual video parsing (AVVP). Each video segment forms a graph node, with event posterior-driven, decay-modulated multi-hop temporal adjacency (Chen et al., 4 Sep 2025).

Edge Weighting:

Edge from node $t$ to $t+l$ weighted by category-dependent hop preference and exponential decay:

$w_{t, t+l} = s_{t,i}\,\exp(-\lambda_t l)$

Two residual GAT layers perform context propagation across the temporal graph.

Performance: On LLP, CATS yields Event@AV of 73.9% (+5.1pp over SOTA), and on UnAV-100, segment-level mAP of 41.9% (Chen et al., 4 Sep 2025).

5. CATS in Multivariate Time Series

a. Auxiliary Time Series Construction for Forecasting

CATS Net constructs multiple auxiliary time series (ATS) as exogenous variables, combined with original series as input to a simple MLP predictor. Key principles—continuity, sparsity, variability—are imposed via separate modules:

• Continuity: Low-pass regularization of ATS.
• Sparsity: Channel- and time-wise gating.
• Variability: Diverse convolutional and projection-based constructors.

The architecture achieves state-of-the-art MTS forecasting accuracy across multiple public datasets, with extremely low parameter counts and computational cost (Lu et al., 2024).

Empirical Summary:

CATS (2-layer MLP core) achieves mean MSE 86.1% relative to DLinear (100%), ranking first in 28/36 forecasting tasks (Lu et al., 2024).

b. Correlation Adapter for Domain-Adaptive Classification

The CATS adapter solves correlation shift in unsupervised domain adaptation for MTS classification (Lin et al., 5 Apr 2025). After each Transformer block, a residual component comprising two temporal depthwise convolution layers and a GAT reweights correlations among variables.

Correlation Alignment Loss:

$$\mathcal{L}_{\mathrm{corr}} = \sum_{k=1}^K \mathrm{MMD}( \operatorname{corr}(H_s^{(k)}), \operatorname{corr}(H_t^{(k)}) )$$

with $$\operatorname{corr}(H) = \operatorname{vec}(H H^{\top} / \|H\|_F^2)$$.

Findings:

Transformers with CATS adapters surpass SOTA baselines by 1–7 percentage points on MTS UDA benchmarks, with only ~1% additional parameters (Lin et al., 5 Apr 2025).

6. Cost Aggregation Transformers for Dense Visual Correspondence

Cost Aggregation Transformers (CATs) target semantic correspondence by refining a 4D cost volume with multi-level, bi-directional (swapping) Transformer aggregation. Appearance affinity modeling and alternating intra/inter-level self-attention operations enable global consensus among pixel matches (Cho et al., 2021).

Architectural Steps:

• Augment cost volume with projected feature embeddings.
• Alternate intra- and inter-correlation map self-attention.
• Swapping self-attention enforces correspondence reciprocity.
• Residual connections stabilize the mapping.

Empirical Results:

• SPair-71k: [email protected] of 49.9% (SOTA).
• PF-PASCAL/WILLOW: 92.6%/90.3% ([email protected]) (Cho et al., 2021).

Table: Major CATS Net Variants and Domains

CATS Variant	Problem Domain	Key Technical Elements
CATS (Huan et al., 2020)	Edge Detection	Tracing loss, pixelwise attention fusion
CATS, CATS v2 (Li et al., 2022, Li et al., 2023)	Medical Segmentation	Dual encoders (CNN+Trans), fusion at skip-connections
CATS (Asavisanu et al., 1 Mar 2025)	V2X Trust/Security	Hybrid majority & reputation consensus, privacy via VPC
CATS (Chen et al., 4 Sep 2025)	Audio-Visual Video Parsing	Category-aware, decay-tempered temporal GAT
CATS (Lu et al., 2024)	MTS Forecasting	ATS construction, 2D contextual attention, simple predictor
CATS Adapter (Lin et al., 5 Apr 2025)	MTS Domain Adaptation	Temporal GAT, correlation alignment loss, minimal overhead
CATs (Cho et al., 2021)	Visual Correspondence	Global Transformer cost aggregation, multi-level, swapping SA

7. Significance and Cross-Domain Impact

The various CATS Net systems introduce distinct algorithmic solutions tailored to core challenges in edge localization, segmentation, temporal reasoning, trust, and time series learning. All share an emphasis on compositionality—fusing complementary representations or integrating multiple consensus mechanisms—and demonstrate efficiency through parameter-light adaptations or attention-based aggregation.

In each target application, CATS Net has been shown to provide either substantial quantitative gains over established baselines or to offer new functionality (such as practical security guarantees in V2X environments) not easily achieved with prior paradigms. This taxonomy highlights the importance of combinatorial module design (fusion, attention, graph construction, consensus) as a central theme across the family of CATS architectures.