SeqNet: Sequential Neural Architectures

• SeqNet is a collection of neural architectures that leverage sequential and hierarchical data processing to enhance pattern recognition across diverse fields such as visual place recognition, person search, biomedical imaging, and malware detection.
• The models use lightweight temporal convolutions, sequential pipelines, and integrated loss functions, yielding improved recall metrics and computational efficiency while handling domain shifts and perceptual aliasing.
• Domain-specific implementations, like SeqNetVLAD and CBGM for person search, demonstrate practical gains in speed and accuracy through strategies like context matching and structured score aggregation.

SeqNet refers to a set of neural network architectures, each independently named "SeqNet," that apply sequential or hierarchical processing to diverse pattern recognition tasks including visual place recognition, person search, biomedical image analysis, and static sequence classification. Although the SeqNet name spans multiple research domains, common features include explicit exploitation of sequential or structured data, computational efficiency, and the integration of temporal or multi-step processes into end-to-end learned pipelines.

1. SeqNet for Sequence-based Visual Place Recognition

SeqNet, introduced by Garg et al. (2021), addresses Visual Place Recognition (VPR) by generating compact learned descriptors that encode short sequences of images, providing robustness against appearance variation (e.g., day–night, seasons) and perceptual aliasing. Key technical contributions are:

• A one-layer temporal convolutional network consumes precomputed single-image descriptors (e.g., NetVLAD, dimension $D=4096$), producing a "sequence descriptor" for an ordered window of $L_d$ frames, $X \in \mathbb{R}^{L_d \times D}$.
• The temporal convolution (kernel width $w$) operates across the sequence, followed by sequence average pooling and L2-normalization:

$$S_{L_d} = \frac{1}{L_d-w+1}\sum_{k=1}^{L_d-w+1} Y_k / \| S̃_{L_d} \|_2$$

• Learned via max-margin triplet loss over sequence descriptors, sampled within-exemplar and across negative spatial radii.
• When $L_d=1$, SeqNet reduces to a linear transformation of a single image descriptor.
• Integrated into a two-stage hierarchical VPR (HVPR) pipeline:
  1. SeqNet generates a shortlist of candidate matches for a given query window.
  2. Selective sequence-based score aggregation reranks these using fast single-image descriptor matching, thus constraining the expensive matching stage to the top-K hypotheses.

Key empirical results: On challenging benchmarks (Oxford RobotCar, Brisbane, Nordland, MSLS), SeqNet's sequence descriptor ($L_d=5$) outperforms single-image NetVLAD by substantial margins (Oxford: Recall@1 of $0.62$ vs. $0.47$; see Table below). The HVPR pipeline matches or exceeds state-of-the-art accuracy while dramatically reducing computation—full-database sequence matching reduces to $O(K L_m)$ computations rather than $L_d$0, with $L_d$1 (Garg et al., 2021).

Method	Oxford R@1	Brisbane R@1
NetVLAD (single)	0.47	0.20
SeqNet $L_d$2	0.62	0.32
SeqNet ($L_d$3+SeqM)	0.73	0.28
HVPR ($L_d$4)	0.72	0.29

This architecture is robust to route reversals (less sensitivity to strict sequence ordering), and with $L_d$5, achieves a good trade-off between latency and discriminative power. Inference runs in under 30 ms per step, suitable for deployment on mobile platforms.

2. SeqNet for Sequential Person Search

In person search, SeqNet (Yan et al., 2021) replaces the typical parallel Faster R-CNN multi-task scheme with explicit sequential sub-networks: detection and re-identification (re-ID) are computed in series rather than in parallel. The first head produces high-quality person detections; only these refined boxes are processed by the second head to generate discriminative embeddings for re-ID.

• Shared ResNet-50 backbone (res1–res4) produces feature maps, feeding two sequential heads.

• Head 1 (detection): Generates object proposals, processes regions (RoIAlign + res5), and produces refined bounding boxes.
• Head 2 (re-ID): Extracts feature vectors from bounding boxes output by Head 1, yielding 256-dimensional embeddings.
• Joint loss across detection (localization and classification) and re-ID (classification + OIM/NAE loss).
• Introduces the Context Bipartite Graph Matching (CBGM) algorithm: a graph-based matching strategy leveraging context people for more robust re-identification by solving a bipartite matching problem via the Hungarian algorithm.

Empirical results demonstrate improved mean Average Precision (CUHK-SYSU: mAP 94.8%, top-1 95.7%) and inference speed (11.5 fps), outperforming parallel head designs and prior person search approaches. The sequential arrangement can be integrated into other end-to-end frameworks, providing immediate gains (Li et al., 2021).

3. SeqNet for Joint Vessel Segmentation and Classification

SeqNet, proposed for retinal vessel analysis in (Li et al., 2020), employs a sequential two-step pipeline:

• Step 1: Vessel segmentation using a UNet/IterNet-style encoder–decoder with iterative refinement, producing a vessel map $L_d$6.
• Step 2: Artery/vein classification using a second UNet applied to the input image masked by the vessel map ($L_d$7), with a "stop-gradient" operation to decouple segmentation and classification training.
• Training is driven by the sum of binary cross-entropy (segmentation) and categorical cross-entropy (classification) losses.
• Structural post-processing: two-stage skeleton-based propagation method unifies intra-vessel labels and propagates high-confidence predictions across segments, using geometric and topological vessel properties for edge normalization.

On the DRIVE dataset, SeqNet achieves vessel segmentation AUC of 0.98, and, after post-processing, centerline artery/vein classification accuracy of 0.919, outperforming or matching state-of-the-art ($L_d$8–$L_d$9\% gain from propagation), with efficient and interpretable label propagation (Li et al., 2020).

4. SeqNet in Malware Detection

The SeqNet model for malware detection is a lightweight, fully 1D convolutional neural network designed for efficient static analysis of raw binary files (Xu et al., 2022):

• Input: Raw binary sequence (≤256 KB), normalized to floats in $X \in \mathbb{R}^{L_d \times D}$0 and up/downsampled by 1D interpolation to fixed length $X \in \mathbb{R}^{L_d \times D}$1.
• Architecture: Embedding Conv1D layer (1→128 channels), followed by five residual Sequence Depthwise-Separable Convolution (Res-SDSC) blocks, global average pooling, and a linear classifier (128→2).
• Total parameter count: ~136K.
• SDSC operation factorizes standard 1D convolution by channel, reducing parameter cost and computation.
• No 2D image conversion, preventing loss of bytewise context and boundary-induced artifacts.

In direct comparison with MalConv and MobileNet-based models, SeqNet achieves $X \in \mathbb{R}^{L_d \times D}$2 accuracy, with 193 MFlops per forward pass and minimal resource footprint. The design ensures robustness against adversarial perturbations at binary boundaries and outperforms 2D convolutional approaches on efficiency with only marginal drops in accuracy (Xu et al., 2022).

Model	Params	Accuracy	Training Time/Epoch
MalConvGCT	1.2M	0.980	2h 38m
SeqNet	136K	0.974	2m 51s
MobileNet	2.2M	0.964	2m 36s

5. SeqNetVLAD: Image Sequence vs. 3D Structure in Place Recognition

Oriol Valls et al. compared the performance of sequence-based image representation (SeqNetVLAD) with point cloud-based approaches (PointNetVLAD) in long-term place recognition (Garg et al., 2021). The key findings are:

• SeqNetVLAD stacks NetVLAD global descriptors from a short sequence of frames, applying two Conv1D+BN+ReLU layers, global temporal pooling, and a final FC+L2 norm to yield a 4096-dimensional sequential descriptor.
• Metric span ($X \in \mathbb{R}^{L_d \times D}$3 frames, $X \in \mathbb{R}^{L_d \times D}$4 m) yields a 10 m effective coverage—shorter than 3D LiDAR methods (PointNetVLAD, 20 m).
• Batch-hard triplet margin loss is used to optimize recall at different top-K levels under day–night domain shift.
• On Oxford RobotCar, SeqNetVLAD with a 10 m span outperforms PointNetVLAD with a 20 m span on Recall@1/5/20 by up to 10 percentage points, demonstrating that short image sequences can match, and even surpass, explicit 3D structure for place recognition under extreme condition change.
• The hierarchical VPR approach further improves localization accuracy by restricting sequence matching to top candidates.

6. Implementation, Runtime, and Practical Considerations

SeqNet architectures are consistently designed for efficiency across domains:

• Visual Place Recognition and SeqNetVLAD run under 30 ms per query on consumer GPUs, using $X \in \mathbb{R}^{L_d \times D}$5, $X \in \mathbb{R}^{L_d \times D}$6, $X \in \mathbb{R}^{L_d \times D}$7 for shortlist size. Memory and computation scale linearly with shortlist size due to the hierarchical pipeline.
• Person search inference achieves 11.5 fps on a single V100 for full detection and re-ID, with CBGM (context matching) adding only 2 ms per query for up to 4,000 gallery images.
• The malware detection SeqNet trains in under 3 minutes per epoch on a single RTX 3090, with a model size below 0.55 MB in parameters, substantially lower than comparators.

Data, models, and code for multiple SeqNet variants are publicly available, supporting direct reproduction and adaptation (Garg et al., 2021, Xu et al., 2022).

7. Conceptual Themes and Domain Adaptation

SeqNet, as realized in various domains, represents a general shift from single-instance static analysis to sequential or structured data modeling, emphasizing:

• Lightweight, shallow temporal or sequential modules over deep recurrent models.
• Flexibility of the core sequential modeling, adapting to applications ranging from structure-aware visual recognition to 1D sequence classification.
• Hierarchical, modular architectures enabling plug-and-play replacement of standard blocks in larger pipelines (notably in person search and VPR).
• Consistent focus on computational tractability, minimal parameterization, and robust performance under domain shift and perceptual aliasing.

Application-specific implementations of SeqNet differ at the instance level, but collectively demonstrate the efficacy of sequential modeling principles when paired with efficient network designs.

References:

• SeqNet for VPR: (Garg et al., 2021)
• SeqNet for person search: (Li et al., 2021)
• SeqNet for vessel segmentation/classification: (Li et al., 2020)
• SeqNetVLAD vs. PointNetVLAD: (Garg et al., 2021)
• SeqNet for malware detection: (Xu et al., 2022)