PPGNet-Cat: Feral Cat Re-ID Model

• The paper introduces PPGNet-Cat, a re-identification model that integrates pose-guided part supervision with global features to achieve a mAP of 0.86 and Rank-1 accuracy of 0.95 on the WA-feral dataset.
• It employs a ResNet-152-based full-image stream alongside modified partial streams for trunk, limbs, and tail, capturing feline morphology under tough conditions.
• The design ensures efficient inference by using only the full-image stream while leveraging manual keypoint annotations during training to overcome low-light, motion blur, and occlusion challenges.

Searching arXiv for the cited paper and closely related PPGNet variants to ground the article. PPGNet-Cat is a feral cat re-identification model for camera-trap imagery that adapts a part-pose guided network originally used in Amur tiger re-ID to the visual and operational constraints of wild-cat monitoring. Its central design combines global appearance features with pose-guided part supervision during training, while restricting inference to a single full-image stream for efficient retrieval. In the reported WA-feral experiments, the model achieved a mean Average Precision (mAP) of 0.86 and a rank-1 accuracy of 0.95, establishing it as a competitive model for feral cat re-ID in the tested setting (Caquilpan, 15 Jul 2025).

1. Task setting and problem definition

PPGNet-Cat is designed for feral cat re-identification: recognizing the same individual cat across images captured by different camera traps, locations, and times. The operational setting is wildlife monitoring with motion-triggered cameras producing Full HD/2K imagery, frequently under nocturnal and low-illumination conditions. The reported difficulties are concrete and domain-specific: low light and high sensor gain at night, small or distant targets, motion blur, partial occlusion by branches or sticks, background clutter from baiting equipment, and limited inter-individual distinctiveness in coat patterns. The paper also notes that black cats often lack distinctive markings, making identity separation especially difficult (Caquilpan, 15 Jul 2025).

The model is motivated by the observation that part-based representations can stabilize identity cues across viewpoint and pose changes. In this setting, pose-guided cropping is used to isolate trunk, limbs, and tail regions so that local descriptors remain more stable than whole-image appearance alone. This is particularly relevant when inter-identity variation is small, as with similar tabby pelage patterns, and when camera-trap images contain large nuisance variation unrelated to the animal itself (Caquilpan, 15 Jul 2025).

The evaluation protocol is retrieval-based rather than closed-set classification in the ordinary sense. Gallery and query images are compared through embedding similarity, and performance is summarized with mAP and cumulative matching characteristics Rank-1. A distinctive aspect of the dataset protocol is that each individual is partitioned into entities by side and time, specifically left/right and day/night, so that appearance changes associated with viewpoint and illumination are made explicit in the label structure (Caquilpan, 15 Jul 2025).

2. Lineage and nomenclatural disambiguation

PPGNet-Cat is a modification of the part-pose guided re-identification architecture referred to in the paper as PPGNet, originally used for Amur tigers by Liu, Zhang, and Guo in 2019. In that lineage, the architecture contains a full-image stream and two partial streams, with the partial streams acting during training as regulators or guidance for the global branch. PPGNet-Cat preserves that high-level structure but changes the part definitions and embedding layout to reflect feline morphology and the properties of feral cat imagery (Caquilpan, 15 Jul 2025).

The name can be confused with several unrelated models. It is not a variant of "P$^2$GNet: Pose-Guided Point Cloud Generating Networks for 6-DoF Object Pose Estimation" (Yu et al., 2019), which is an RGB-D 6-DoF pose estimator for known object instances. It is also unrelated to "Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space" (Nguyen et al., 2016), where “PPGN-class” is sometimes colloquially described as a category-conditioned or “cat” variant in the sense of class-conditional generation. Likewise, it is distinct from "PPGnet: Deep Network for Device Independent Heart Rate Estimation from Photoplethysmogram" (A et al., 2019) and from "PPGNet: Learning Point-Pair Graph for Line Segment Detection" (Zhang et al., 2019). In short, PPGNet-Cat is an animal re-ID model, not a pose-estimation, generative-image, physiological-signal, or line-segment-detection system.

This disambiguation matters because the term “PPGNet” has been used for multiple architectures across CV and signal processing. Within the feral-cat work, however, PPGNet-Cat refers specifically to the cat-adapted, part-pose guided re-ID model introduced for wildlife monitoring (Caquilpan, 15 Jul 2025).

3. Architecture and cat-specific modifications

The original PPGNet structure comprises three streams. The F-Stream is a full-image branch for global features. The TP-Stream is a trunk partial stream for local trunk features. The LP-Stream is a limb partial stream for local limb features. In the original formulation, TP-Stream and LP-Stream guide F-Stream during training, but only F-Stream is used at inference, which keeps test-time prediction lighter and faster (Caquilpan, 15 Jul 2025).

PPGNet-Cat retains this training/inference asymmetry. It uses ResNet-152 for F-Stream and ResNet-34 for the partial streams, with pretrained weights; the paper states that ImageNet pretraining is implied by the Batch Normalization comment, but the exact source is not explicitly stated. Batch Normalization is present and becomes relevant for a specific engineering modification: when part crops are absent and would otherwise produce black images, the corresponding partial-stream embeddings are set to zero to avoid BN-induced noise (Caquilpan, 15 Jul 2025).

The cat-specific changes are centered on pose definition and part extraction. The tiger ATRW 15-keypoint scheme was replicated by manually annotating cat images where possible, and two additional cat-specific keypoints were added for the proximal and distal tail. No automatic cat pose-estimation model was trained; keypoints were manually annotated in the training images, and no confidence scores or visibility estimation are used. This yields three notable geometric changes relative to the tiger setting:

• Limbs: rectangular crops are defined with width set to one-third of height, i.e., height/width $= 3$, to reflect thinner cat limbs.
• Trunk: the original axis-aligned crop is replaced by a rotated rectangle aligned to body orientation, addressing frequent non-horizontal poses.
• Tail: two additional regions, proximal and distal tail, are cropped using the added tail keypoints, with extraction following the limb strategy.

The embedding structure is also modified. Two tail embeddings of size 256 each are added in the LP-Stream, described as “half the size of the embeddings for the limbs.” The final embedding dimensionality is increased from 2,048 in the original PPGNet to 2,560 in PPGNet-Cat. During training, the paper states that trunk and limb features are “summed” to the F-Stream embedding to create guided features, but the exact fusion operator and ordering are not exhaustively specified. At inference, only the F-Stream 2,560-D embedding is used (Caquilpan, 15 Jul 2025).

A common misconception is that the pose-guided machinery must remain active at test time. In fact, inference does not require keypoints; the pose-guided part streams are a training-time supervisory mechanism, while deployment uses only the global stream (Caquilpan, 15 Jul 2025).

4. Datasets, annotation protocol, and optimization setting

The methodological baseline is linked to ATRW, the Amur Tiger Re-ID in the Wild dataset, which contains 3,649 images, 92 identities, and 15 annotated keypoints. For the feral-cat study, the primary dataset is WA-feral, with 3,120 images of 10 distinct cats collected from multiple camera traps and locations in Western Australia. A separate Victoria-feral dataset contains 21 cats across 35 entities and is used only for cross-dataset inference and evaluation (Caquilpan, 15 Jul 2025).

The preprocessing pipeline on WA-feral removes sequences of near-identical frames to reduce overfitting, excludes images showing only partial bodies such as just a head or tail, and applies YOLOv8 for bounding-box cropping. The identity protocol then partitions each cat into entities by side and time, namely left/right and day/night. The paper emphasizes that this partitioning is crucial for performance. The reported train/test split on WA-feral is 60/40 on entities, yielding 20 train entities with 424 images and 10 test entities with 328 images; test entities are disjoint from train, mirroring ATRW protocol (Caquilpan, 15 Jul 2025).

Training uses PyTorch on an NVIDIA GeForce RTX 2080 Ti, typically for 100–300 epochs. The optimizer, learning rate, schedule, batch size, and weight decay are not specified. Augmentation includes Gaussian blur, Gaussian noise injection, random erasing, perspective transformation, and rotation. These augmentations are intended to simulate low-light degradation, occlusion, camera-angle changes, and orientation variability (Caquilpan, 15 Jul 2025).

The supervised objectives are ID loss and triplet loss on embeddings, while ArcFace was explored in some experiments. The exact mathematical forms and overall weighting are not provided in the paper. The paper reports that ArcFace improved a simple ResNet-152 baseline and an adapted PPGNet* baseline, but did not improve PPGNet-Cat when trained to convergence; some short-epoch runs under 100 epochs showed transient improvements, whereas longer training made ArcFace ineffective for PPGNet-Cat in the reported setup (Caquilpan, 15 Jul 2025).

5. Empirical performance and ablation structure

The principal WA-feral results compare a plain ResNet-152 baseline, the same baseline with ArcFace, an adapted PPGNet* baseline, PPGNet* with ArcFace, PPGNet-Cat, and PPGNet-Cat with ArcFace. The reported values are as follows (Caquilpan, 15 Jul 2025):

Model	mAP	Rank-1
ResNet-152 baseline	0.44	0.62
ResNet-152 + ArcFace	0.60	0.83
PPGNet*	0.77	0.91
PPGNet* + ArcFace	0.82	0.95
PPGNet-Cat	0.86	0.95
PPGNet-Cat + ArcFace	0.82	0.94

These results place PPGNet-Cat above the reported baselines in mAP while matching the best reported Rank-1. They also show that ArcFace is not uniformly beneficial: it improves weaker baselines but reduces the best reported PPGNet-Cat configuration after convergence. A plausible implication is that the cat-specific architectural changes already absorb much of the discriminative gain that ArcFace provides to simpler embeddings, although the paper does not formalize that interpretation (Caquilpan, 15 Jul 2025).

The ablation sequence isolates the contribution of specific design choices. Base PPGNet is reported at approximately 0.75 mAP and approximately 0.89 Rank-1. Adding random erasing raises performance to approximately 0.82 mAP and approximately 0.93 Rank-1. Forcing zeros for black part-crops gives approximately 0.825 mAP and approximately 0.95 Rank-1, which the paper interprets as reducing BN noise. Adding the rotated trunk crop yields approximately 0.83 mAP and approximately 0.95 Rank-1. Adding the tail, which defines PPGNet-Cat, yields approximately 0.86 mAP and approximately 0.95 Rank-1 (Caquilpan, 15 Jul 2025).

Partitioning by side and time is one of the most consequential protocol choices. Without partitioning, performance is reported as mAP 0.73 and Rank-1 0.93. Day/night only gives mAP 0.78 and Rank-1 0.91. Left/right only gives mAP 0.48 and Rank-1 0.81. The combined left/right plus day/night partition gives the best result, mAP 0.86 and Rank-1 0.95. This indicates that side alone harms retrieval, whereas side jointly modeled with illumination regime becomes beneficial in the reported data organization (Caquilpan, 15 Jul 2025).

Cross-dataset generalization to Victoria-feral drops to mAP 0.73 and Rank-1 0.87. The paper states that this is plausibly due to more individuals, specifically 21 cats across 35 entities, and domain differences. UMAP visualization of the 2,560-D embeddings shows strong clustering by entity, with two prominent groups corresponding to left/right nocturnal entities and a small day cluster, the latter reflecting that only 10 day images appear in the test set. Error analysis identifies near-identical tabby patterning, background artifacts such as a white baiting stick, blurry low-light imagery, and uniformly black cats as salient failure modes (Caquilpan, 15 Jul 2025).

6. Inference pipeline, system footprint, and limitations

The inference pipeline is intentionally simpler than the training graph. First, YOLOv8 detects the cat bounding box and the image is cropped. Second, the cropped image is passed through the F-Stream, implemented with ResNet-152, to produce a 2,560-D embedding. Third, this embedding is compared with gallery embeddings using cosine similarity or cosine distance. The paper gives the distance explicitly as

$$d_{\cos}(x, y) = 1 - \frac{x^\top y}{\|x\|\,\|y\|}.$$

Gallery items are ranked by ascending $d_{\cos}$, and retrieval metrics such as Rank-1 and mAP are then computed over the queries (Caquilpan, 15 Jul 2025).

The training-time model is substantially larger than the deployed model. The reported training model size is approximately 197M parameters, or approximately 789 MB. The inference model, which retains only F-Stream, is approximately 70M parameters, or approximately 283 MB. Detailed throughput, FPS, and edge-device memory behavior are not reported, but the single-stream inference design is presented as the efficiency mechanism that makes deployment more tractable than multi-stream evaluation (Caquilpan, 15 Jul 2025).

Several limitations are explicit. Training depends on manual keypoint annotation because no automatic cat pose estimator is used. Domain shift remains significant, as shown by the Victoria-feral performance drop. Low-SNR night images remain challenging, especially because the pose estimator is not employed at inference. Small cats, kittens, and distant targets reduce feature quality through limited resolution. Daytime images are underrepresented in training, and long-term appearance changes were not studied. Strong background artifacts may spuriously drive similarity, particularly when they recur in the camera-trap environment (Caquilpan, 15 Jul 2025).

The paper identifies two immediate directions for future work: increasing training dataset size and diversity, and adjusting or fine-tuning hyperparameters that were largely inherited from the tiger PPGNet context. These directions follow directly from the observed dependence on annotation quality, data partitioning, and the remaining sensitivity to cross-dataset variation (Caquilpan, 15 Jul 2025).