TACO-Net: 3D Topological Object Classification

• The paper demonstrates that TACO-Net converts point clouds into voxelized 3D images and extracts topological signatures via persistent homology for robust object classification.
• It employs multiple filtrations—such as height, radial, and density-based—to generate a compact 36-dimensional feature vector per filtration, ultimately classified by a lightweight 1D CNN.
• Empirical results on CAD and real-world datasets highlight state-of-the-art accuracy and impressive robustness under noise, despite the relatively high computational cost of feature extraction.

Searching arXiv for the exact TACO-Net paper and related disambiguation papers. TACO-Net is a 3D object classification framework for point clouds that converts each input point cloud into a voxelized binary 3D image, applies multiple grayscale filtrations, computes cubical persistent homology on the resulting cubical complexes, converts each persistence output into a compact topological descriptor, and classifies the concatenated descriptor with a lightweight 1D CNN. In the paper "TACO-Net: Topological Signatures Triumph in 3D Object Classification" (Ghosh et al., 29 Sep 2025), the method is presented as a topological data analysis-based alternative to point-based, graph-based, transformer-based, and multi-view pipelines, with the stated goal of improving robustness to unordered, irregular, and noisy point-cloud representations.

1. Name, scope, and disambiguation

The exact method name is TACO-Net, expanded in the paper as "Topological data analysis-based object classification" (Ghosh et al., 29 Sep 2025). Its scope is specifically 3D object classification from point clouds. The pipeline is therefore not a generic neural architecture for arbitrary modalities, but a task-specific system in which topological summaries derived from voxelized 3D shapes are the primary representation.

This usage should be distinguished from several similarly named arXiv works. In particular, "TACNet: Transition-Aware Context Network for Spatio-Temporal Action Detection" (Song et al., 2019) is a video action-detection model whose official name is TACNet, not TACO-Net. The supplied corpus also contains multiple papers titled TACO or TaCo in unrelated areas such as tensor-parallel LLM training, continual neural mapping, remote sensing change detection, semantic communications, tabular foundation models, terminal-agent compression, tool-use RL, column description generation, and quadrotor control; those papers explicitly use TACO or TaCo, not TACO-Net. In the strict bibliographic sense, the name TACO-Net in this material refers to the 3D object-classification system of (Ghosh et al., 29 Sep 2025).

2. Task setting and motivation

TACO-Net addresses 3D object classification, where the input is a point cloud sampled either from CAD models or from real sensing pipelines such as LiDAR or RGBD acquisition, and the output is a semantic class label. The paper emphasizes that point-cloud classification remains difficult because point clouds are unordered, irregular and sparse, often affected by noise, occlusion, and varying point density, and lack a regular image-like grid while still requiring the model to capture both local and global geometry (Ghosh et al., 29 Sep 2025).

The method’s central thesis is that directly learning from raw point sets is not the only viable route. Instead, TACO-Net treats the point cloud as a source from which one can derive topological signatures. The authors argue that topology captures global shape structure and can be more robust to perturbations than methods that depend heavily on exact point coordinates. This positioning is reinforced by the paper’s emphasis on persistent homology, especially long-lived topological features, as descriptors of structural organization rather than incidental sampling noise.

The claimed contribution is twofold at the algorithmic level. First, the framework uses cubical persistence on point-cloud-derived voxel data before learning with a 1D CNN. Second, it combines several complementary filtrations so that connected components, tunnels, cavities, local density effects, and boundary-distance effects all contribute to the final descriptor. A plausible implication is that the method trades end-to-end differentiability for a highly structured, geometry-driven representation pipeline.

3. Representation pipeline and multi-filtration topology

The end-to-end pipeline begins by transforming each train and test point cloud $P$ into a 3D binary image $\mathcal{B}$, in which every voxel containing at least one point is activated and every other voxel is deactivated (Ghosh et al., 29 Sep 2025). The formal definition given is

$\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$

where each $p \in I$ is a voxel. The paper states that the volume of the binary image is roughly equal to the axis-parallel bounding box of the point cloud. For ModelNet40 and ModelNet10, the best voxel size reported is 0.05.

TACO-Net then maps the binary image to multiple grayscale 3D images

$$\mathcal{G}: I \subseteq \mathbf{Z}^3 \to \mathbf{R},$$

using six filtration families: height filtration, radial filtration, density filtration, dilation filtration, erosion filtration, and signed distance filtration. These filtrations are intended to expose different structural aspects of the same object. The paper particularly stresses that the 26 directional height filtrations over the nonzero directions in $\{0,1,-1\}^3\setminus\{(0,0,0)\}$ discretize the Persistent Homology Transform, while radial filtrations contribute information about interior organization.

The full design uses 26 height filtrations, 27 radial filtrations, and one instance each of density, dilation, erosion, and signed distance, for a total of

$26 + 27 + 4 = 57$

filtrations. For ModelNet40 and ModelNet10, however, the best configuration reported uses 26 height, 18 radial, and the four density/erosion/dilation/signed-distance filtrations, giving

$26 + 18 + 4 = 48$

filtrations. This reduction is tied to the paper’s feature-length ablation.

Each grayscale image is interpreted as a cubical complex and processed through a sublevel-set filtration

$K_0 \subseteq K_1 \subseteq \cdots \subseteq K_m.$

Persistent homology is then computed in dimensions $H_0$, $\mathcal{B}$0, and $\mathcal{B}$1, corresponding respectively to connected components, tunnels or loops, and enclosed cavities or voids. The resulting cubical persistence is represented as a multiset

$\mathcal{B}$2

which constitutes the topological signature for one filtration of one object (Ghosh et al., 29 Sep 2025).

4. Persistent-homology summaries and feature construction

TACO-Net does not use raw persistence diagrams directly as classifier input. Instead, it converts each filtration’s persistence output into a 36-dimensional feature vector (Ghosh et al., 29 Sep 2025). The first three dimensions are the persistent entropies of $\mathcal{B}$3, $\mathcal{B}$4, and $\mathcal{B}$5, defined by

$\mathcal{B}$6

The remaining 33 dimensions are amplitudes derived from several standard persistence summaries. The Wasserstein amplitudes are

$\mathcal{B}$7

used with $\mathcal{B}$8, producing six features across the three homology dimensions. The Bottleneck amplitude is

$\mathcal{B}$9

producing three more features. The Betti-curve amplitudes contribute six features via $\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$0 and $\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$1 norms over three homology dimensions, using 100 filtration values. The persistence landscape amplitudes contribute twelve features, with landscapes defined from

$\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$2

using $\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$3 and $\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$4, $\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$5 norms, again across the three homology dimensions. Finally, the heat-kernel amplitudes contribute six features, with $\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$6 and 20 filtration values.

The per-filtration dimensionality is therefore

$\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$7

If all 57 filtrations are retained, the descriptor length is

$\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$8

In the best ModelNet40/10 configuration, the final feature length is

$\mathcal{B}: I \subseteq \mathbf{Z}^3 \to \{0,1\},$9

This descriptor is then fed to a lightweight 1D CNN. The paper specifies a stack of three 1D convolution layers with kernel sizes $p \in I$0, $p \in I$1, and $p \in I$2, and filter counts 128, 64, and then the number of classes for ModelNet40/10. Each convolution is followed by BatchNorm and ReLU, after which a fully connected layer of size equal to the number of classes and a softmax produce the final class prediction (Ghosh et al., 29 Sep 2025). The paper does not specify stride, padding, pooling, dropout, or residual connections.

5. Training configuration, implementation, and complexity

For ModelNet40 and ModelNet10, each object is represented by a random uniform sample of 2048 3D points before voxelization (Ghosh et al., 29 Sep 2025). The training setup uses the Adam optimizer with learning rate 0.001, minibatch size 128, and maximum 1000 training epochs. Early stopping is triggered when the training loss reaches 0.005, and each configuration is trained 5 times, with average results reported unless otherwise specified.

The implementation is split across two environments. The topological pipeline is implemented in Python using giotto-tda, while the CNN is implemented in MATLAB. Hardware reported in the paper consists of an Intel i9-12900K, 32 GB RAM, and an NVIDIA RTX 3060 GPU (Ghosh et al., 29 Sep 2025). The number of learnable parameters is correspondingly small: 0.72M on ModelNet40, 0.71M on ModelNet10, 0.50M on VesselMNIST, and 0.66M on AdrenalMNIST.

The computational profile is asymmetric. Topological feature extraction is the expensive stage, whereas classification after feature extraction is very fast. For ModelNet40 with voxel size $p \in I$3 and feature length 1728, the paper reports 5.3 point clouds/sec for TDA feature generation, versus 16,454 point clouds/sec for classification throughput, with total training time of 2.50 minutes (Ghosh et al., 29 Sep 2025). Voxel size affects this strongly: $p \in I$4 yields 8.2 point clouds/sec, whereas $p \in I$5 yields 1.4 point clouds/sec. The paper also states an end-to-end classification-time complexity of

$p \in I$6

where $p \in I$7 is the number of input points, $p \in I$8 is the number of voxels in $p \in I$9, and $$\mathcal{G}: I \subseteq \mathbf{Z}^3 \to \mathbf{R},$$0 is voxel size.

6. Empirical results, robustness, and ablations

The paper reports state-of-the-art accuracy on the synthetic CAD benchmarks ModelNet40 and ModelNet10, and also evaluates on OmniObject3D, ScanObjectNN, VesselMNIST3D, and AdrenalMNIST3D (Ghosh et al., 29 Sep 2025).

Dataset	Metric(s)	TACO-Net
ModelNet40	OA / mAcc	99.05 / 97.97
ModelNet10	OA / mAcc	99.52 / 99.52
OmniObject3D	OA	58.90
ScanObjectNN	OA	93.94
VesselMNIST3D	mAcc / F1	95.28 / 0.94
AdrenalMNIST3D	OA	80.54

On ModelNet40, the reported 99.05% OA exceeds the strongest baseline listed in the paper, RotationNet at 97.37% OA. On ModelNet10, TACO-Net reaches 99.52% OA, above RotationNet’s 98.9%. On OmniObject3D, the model achieves 58.90% OA, improving over the strongest listed baseline, CurveNet at 50.0%, by 8.9%. On ScanObjectNN, the reported 93.94% OA is described by the authors as not state of the art, but still higher than many standard baselines. The medical-dataset results extend the method beyond object-classification benchmarks: 95.28% mAcc and 0.94 F1 on VesselMNIST3D, and 80.54% OA on AdrenalMNIST3D.

Robustness is evaluated on ten corruption types derived from ModelNet40-style inputs. For low-severity corruptions, the paper states that TACO-Net achieves $$\mathcal{G}: I \subseteq \mathbf{Z}^3 \to \mathbf{R},$$1 OA for all tested perturbations except impulse noise. Clean accuracy for the best model is 99.15 OA and 98.37 mAcc. Under high-severity corruptions, the average OA is 68.65%, with particularly weak robustness to impulse noise: 52.88 OA at low severity and 24.19 OA at high severity. The corruption study therefore supports robustness to many perturbations, but not uniformly across all corruption families.

The ablation studies clarify which ingredients matter most. Using the same CNN on ModelNet40, DEDS only yields 96.52 OA, H only yields 98.29 OA, DEDS + H only yields 98.82 OA, and Entropy only yields 96.16 OA. This indicates that directional height filtrations are the strongest single family, while density/erosion/dilation/signed-distance cues add incremental value and radial filtrations further improve the full model. A depth ablation shows that reducing the CNN to the first two Conv1D layers lowers OA to 98.18, and using only the first Conv1D layer lowers it to 94.76, so all three convolution stages contribute materially. The best voxel size on ModelNet40 is 0.05, and the best feature length is 1728, corresponding to 18 radial centers plus 26 height and 4 DEDS filtrations (Ghosh et al., 29 Sep 2025).

7. Interpretation, limitations, and place within 3D classification

TACO-Net’s most distinctive feature is its reliance on persistent homology as the primary representation. The method uses $$\mathcal{G}: I \subseteq \mathbf{Z}^3 \to \mathbf{R},$$2, $$\mathcal{G}: I \subseteq \mathbf{Z}^3 \to \mathbf{R},$$3, and $$\mathcal{G}: I \subseteq \mathbf{Z}^3 \to \mathbf{R},$$4 explicitly, so its descriptors encode mergers of connected regions, loop or tunnel structure, and enclosed cavities. The paper’s interpretation is that such signatures emphasize global shape organization and the long-lived topological structures that tend to survive perturbation, rather than pointwise geometric detail alone (Ghosh et al., 29 Sep 2025). This helps explain why a compact 1D CNN can perform strongly once the topological descriptors have been extracted.

The method also has clear tradeoffs. The pipeline is not end-to-end differentiable through persistent homology in the paper’s implementation, and the expensive step is the preprocessing stage rather than the classifier. The authors explicitly identify feature generation as the main bottleneck and suggest future acceleration through GPU parallelism. Robustness is strong overall but not universal: severe corruptions that substantially alter the underlying shape remain difficult, and impulse noise is the clearest weakness reported. The paper also notes that while TACO-Net performs strongly on real-world datasets such as OmniObject3D and competitively on ScanObjectNN, it does not directly validate sparse-LiDAR classification despite mentioning LiDAR in motivation.

Within the 3D object-classification literature, TACO-Net can therefore be understood as a structured alternative to raw-point or transformer-heavy pipelines: it first converts a point cloud into a voxelized volume, then into a family of filtrations, then into persistence-derived numerical summaries, and only then applies a small neural classifier. This suggests a broader methodological point: for some 3D recognition problems, strong performance can arise from carefully engineered topological signatures coupled with a lightweight learner, rather than from increasingly large point-cloud backbones alone.