---
title: 'Pi3DET: 3D a-Si:H Detectors & LiDAR Detection'
url: https://www.emergentmind.com/topics/pi3det
type: topic
---

# Pi3DET: 3D a-Si:H Detectors & LiDAR Detection

Searching arXiv for Pi3DET and directly related papers to ground the article.
arxiv_search.query({"search_query":"all:Pi3DET OR ti:\"Perspective-Invariant 3D Object Detection\"","start":0,"max_results":10})
arxiv_search.query({"search_query":"ti:\"Testing of planar hydrogenated amorphous silicon sensors with charge selective contacts for the construction of 3D- detectors\" OR ti:\"Perspective-Invariant 3D Object Detection\"","start":0,"max_results":10})
Pi3DET is a context-dependent term used in the cited literature in two distinct ways. In detector-development work, it denotes a program aimed at building **3D hydrogenated amorphous silicon (a-Si:H) detectors** for radiation detection and dosimetry, with current results centered on planar validation of contact technologies needed for a future 3D architecture [2109.15010]. In autonomous-perception research, **Pi3DET** is introduced as **“Perspective-invariant 3D object DETection”**, the first publicly available **multi-platform LiDAR 3D object detection dataset and benchmark** designed for **cross-platform** detection and adaptation across **vehicle**, **quadruped**, and **drone** platforms [2507.17665]. The coexistence of these usages makes the term domain-specific rather than universally fixed.

## 1. Nomenclature and domain-specific usage

Within the cited arXiv literature, Pi3DET does not designate a single universally adopted object. One usage is a detector-fabrication program focused on **3D a-Si:H detectors**; the other is a LiDAR benchmark and adaptation framework for **perspective-invariant 3D object detection** [2109.15010], [2507.17665]. A related review on **3D silicon detectors** does not introduce Pi3DET as a separate acronym or new detector concept; in that context, if the label is used, it most naturally refers to the broader **3D pixel detector technology** rather than to a distinct device class [1511.02080].

| Usage of “Pi3DET” | Research area | Defining source |
|---|---|---|
| 3D hydrogenated amorphous silicon detector program | Radiation detection and dosimetry | [2109.15010] |
| Perspective-invariant 3D object DETection | LiDAR 3D detection and cross-platform adaptation | [2507.17665] |

A recurrent source of confusion is the assumption that the acronym has a single canonical meaning across detector physics and machine perception. The cited literature does not support that assumption. Instead, meaning is fixed by disciplinary context.

## 2. Pi3DET as a 3D a-Si:H detector program

In the detector-physics usage, Pi3DET is aimed at building **3D hydrogenated amorphous silicon (a-Si:H) detectors** for radiation detection and dosimetry [2109.15010]. The intended 3D geometry differs from conventional planar devices: it would start from an a-Si:H substrate and then introduce **holes or trenches through the bulk**, so that detector electrodes can be formed on the internal walls of those cavities. The attraction of this geometry is that it can **shorten carrier drift distances** and improve **radiation tolerance** and **charge-collection performance** in harsh environments.

The program is motivated by material properties of **hydrogenated amorphous silicon** that are explicitly emphasized in the cited work. a-Si:H is described as a material with **intrinsic radiation hardness** and a history of use in **solar cells** as well as in **particle detection and dosimetry**. A plausible implication is that Pi3DET treats a-Si:H not merely as a substitute for crystalline silicon, but as a thin-film platform whose deposition flexibility and contact engineering may be compatible with unconventional 3D geometries.

The central fabrication difficulty is conformality. The standard planar PECVD approach used to make doped a-Si:H layers is described as **not conformal enough to coat the sidewalls of deep etched structures**. To address that problem, the project considers two junction-fabrication routes:

- **Ion implantation**: phosphorus implantation for **n-type** regions and boron implantation for **p-type** regions.
- **Charge-selective contacts via ALD**: **AZO** or **TiO\(_2\)** as electron-selective contact, and **MoO\(_x\)** as hole-selective contact.

The work reported in the cited paper does not yet realize the final 3D detector. Instead, it evaluates one enabling idea—**charge-selective contacts**—in planar prototypes, so that leakage current, temperature dependence, and x-ray response can be measured in a controlled configuration before attempting integration on 3D sidewalls [2109.15010].

## 3. Planar validation of charge-selective contacts for future 3D integration

The planar prototypes used to validate the Pi3DET contact concept were fabricated on **glass substrates** with an **a-Si:H active layer** of either **6.2 µm** or **8.2 µm** thickness [2109.15010]. The electron-selective contact was either **AZO** of **60 nm** or **TiO\(_2\)** of **10 nm**. The hole-selective contact was **20 nm MoO\(_x\)** protected by a **60 nm ITO** layer. The a-Si:H was deposited by **PECVD** from **silane (SiH\(_4\))** and **molecular hydrogen**, and the hole-selective contacts were patterned into **four square electrodes of 5 × 5 mm\(^2\)** area.

Electrical characterization in a climatic chamber at about **21 °C** showed that **AZO devices leaked less than TiO\(_2\) devices**, and that **thinner 6.2 µm devices had lower leakage current than 8.2 µm devices** at the same field. At **5 V/µm**, the normalized leakage current densities were reported as **9.2 nA/cm\(^2\)** and **14.8 nA/cm\(^2\)** for two **8.2 µm AZO** samples, **1.76 nA/cm\(^2\)** for the **6.2 µm AZO** sample, and **24.4 nA/cm\(^2\)** for the **8.2 µm TiO\(_2\)** sample. Below about **10 °C**, the current dropped below **100 pA** but became unstable. The temperature dependence was described as exponential, for example
$$
y = 0.10228 \, e^{(0.082338x)} \quad (R=0.99973),
$$
with \(y\) the measured leakage current, \(x\) temperature, and \(R\) the regression coefficient. The authors interpret this exponential behavior as evidence that carrier mobility increases with temperature in a-Si:H and infer that **AZO provides better transport behavior than TiO\(_2\)**.

Under x-ray irradiation from a miniature x-ray tube operating up to **50 kV** and **200 µA** with a **125 µm Be window**, and over a dose-rate range of **0–22 mGy/s**, the induced current was defined as the current under illumination minus the dark leakage current [2109.15010]. At **0 V bias**, corresponding to photovoltaic-mode operation relevant to self-powered dosimetry, the linear dose-response sensitivities were:

| Device | 0 V sensitivity | Biased sensitivity |
|---|---:|---:|
| 8.2 µm AZO sample 1 | 1.37 nC/cGy | 16.71 nC/cGy |
| 8.2 µm AZO sample 2 | 1.48 nC/cGy | 17.98 nC/cGy |
| 6.2 µm AZO sample | 0.84 nC/cGy | 8.78 nC/cGy |
| 8.2 µm TiO\(_2\) sample | 1.25 nC/cGy | 16.86 nC/cGy |

The reported regression coefficients at **0 V** were **0.999880**, **0.999975**, **0.999897**, and **0.999918**; under reverse bias they remained **0.999655**, **0.999491**, **0.999828**, and **0.999757**. The applied reverse biases were **30 V** for the **8.2 µm** samples and **20 V** for the **6.2 µm** sample. The cited work states that biasing increased sensitivity by more than a factor of 10 relative to **0 V** operation, bringing the response into the same general performance range as commercial crystalline silicon dosimeters.

Bias dependence at fixed x-ray conditions (**100 µA** tube current, **30 kV**, nominal dose rate **6.8 mGy/s**) was fit with
$$
I = p_0 \left(1 - p_1 e^{-p_2 V}\right),
$$
where \(I\) is photocurrent, \(V\) is bias voltage, and \(p_0, p_1, p_2\) are fit parameters. The saturation photocurrent is approximately \(p_0\), and the voltage for near-saturation is on the order of \(\sim 3/p_2\). Reported saturation voltages and currents were **209.2 V**, **38.48 nA**; **333.4 V**, **64.44 nA**; **144.2 V**, **22.74 nA**; and **199.7 V**, **37.33 nA** for the two **8.2 µm AZO** samples, the **6.2 µm AZO** sample, and the **8.2 µm TiO\(_2\)** sample, respectively.

The stated conclusion is that **charge-selective contacts are feasible and promising for detector use**, especially **AZO** as the electron-selective contact. The cited paper therefore recommends **AZO** as the baseline electron-selective contact material for future **3D a-Si:H detector fabrication**. It also explicitly notes the present limits: **0 V sensitivity** remains lower than crystalline silicon detectors, measurements become unstable below about **10 °C**, performance depends noticeably on **a-Si:H thickness**, and **TiO\(_2\)** underperformed **AZO** in leakage and transport quality [2109.15010].

## 4. Relation to established 3D detector technology

The Pi3DET detector program sits within a broader technological lineage of **3D detectors**, especially **3D silicon detectors**, in which electrodes penetrate the sensor bulk perpendicular to the surface [1511.02080]. In those devices, the electrodes are **columnar doped structures etched into the silicon bulk**, and the crucial design consequence is that the **inter-electrode distance** is decoupled from the **sensor thickness**. The cited review identifies the resulting operational benefits as **lower depletion voltage**, **shorter carrier drift distance**, **faster charge collection**, **less trapping after irradiation**, **high signal even after heavy radiation exposure**, and **natural slim or active edges**.

Fabrication in silicon relies on **deep reactive ion etching (DRIE)** to form the columns, followed by alternating **n\(^+\)** junction and **p\(^+\)** ohmic doping. The review distinguishes **single-sided** and **double-sided** processes, and reports mature deployment in the **ATLAS Insertable B-Layer (IBL)**, together with ongoing work for **AFP**, **PPS**, and the **High-Luminosity LHC** [1511.02080]. Quantitative reference points include **inter-electrode spacing** of about **70 μm** in contemporary devices, future aims near **30 μm**, and present thicknesses around **200 μm**.

This background matters for Pi3DET because the a-Si:H program adopts the same structural intuition—short drift paths enabled by internal electrode geometries—while confronting a distinct materials problem, namely how to realize sidewall junctions in amorphous silicon films. A plausible implication is that Pi3DET in the detector-physics sense can be read as an attempt to transplant the 3D-detector logic of silicon tracking devices into a thin-film **a-Si:H** platform more directly suited to radiation detection and dosimetry.

The silicon review also provides a benchmark for why 3D geometries are attractive in harsh environments. For HL-LHC innermost layers, fluences up to
$$
2\times10^{16}\,\text{n}_{eq}/\text{cm}^2
$$
are cited, together with low-voltage and low-power operation as decisive strengths [1511.02080]. Pi3DET’s emphasis on **intrinsic radiation hardness** of a-Si:H and on improved charge collection in harsh environments is therefore aligned with a broader 3D-detector design philosophy, even though the material system and fabrication route differ.

## 5. Pi3DET as “Perspective-invariant 3D object DETection”

In machine perception, **Pi3DET** is the name of a dataset, benchmark, and associated adaptation framework for LiDAR-based 3D object detection across heterogeneous robot platforms [2507.17665]. The paper introduces it as the first benchmark featuring LiDAR data and 3D bounding box annotations collected from **vehicle**, **quadruped**, and **drone** platforms, specifically to study **cross-platform 3D detection** rather than only conventional cross-dataset transfer.

The dataset is built on **M3ED** and contains **25 sequences**, **51,545 LiDAR frames**, **over 250,000 3D bounding boxes**, and **10 Hz annotations**. Platform-level statistics are reported as follows: **Vehicle** has **8 sequences**, **32,193 frames**, **346.95M points**, **131,911 vehicle boxes**, and **88,986 pedestrian boxes**; **Drone** has **7 sequences**, **7,052 frames**, **59.47M points**, **14,534 vehicle boxes**, and **1,272 pedestrian boxes**; **Quadruped** has **10 sequences**, **12,300 frames**, **156.75M points**, **5,982 vehicle boxes**, and **14,551 pedestrian boxes**. The full dataset contains **563.17 million LiDAR points**. The class set is deliberately restricted to **Vehicle** and **Pedestrian**.

All three platforms use a **64-beam LiDAR** with angular range \([-22.5^\circ, 22.5^\circ]\). The dataset is explicitly designed to expose differences in **point-cloud elevation distribution**, **ego-motion distribution**, and **target box geometry distribution**. The cited paper argues that these shifts are sufficiently strong that detectors trained on vehicle-mounted LiDAR do not transfer well to drones or quadrupeds.

The annotation format uses oriented 3D boxes
$$
\mathbf{b} = (c^{x}, c^{y}, c^{z}, l, w, h, \varphi) \in \mathbb{R}^7,
$$
where \(\mathbf{c}=(c^x,c^y,c^z)\) is the box center, \((l,w,h)\) are box dimensions, and \(\varphi\) is the heading angle. Annotation proceeds through a three-stage pipeline: **pseudo-label generation** using pretrained detectors including **PV-RCNN**, **PV-RCNN++**, **Voxel-RCNN**, **IA-SSD**, **CenterPoint**, and **SECOND**; **pseudo-label optimization and filtering** using **kernel density estimation (KDE)**, **CTRL** multi-object tracking, and **Tokenize Anything (TA)**; and **manual refinement** in **Xtreme1** by **three annotators**, with a total effort of **about 500 hours** and average manual intervention of **over 30 seconds per frame** [2507.17665].

Train/validation splits are reported as **16,888 / 15,305** for **Vehicle**, **7,204 / 5,096** for **Quadruped**, and **3,584 / 3,468** for **Drone**. The benchmark includes **six cross-platform adaptation benchmarks** and **two cross-dataset benchmarks**, for a total of **eight** adaptation tasks. The six cross-platform tasks are:

- **Pi3DET (Vehicle) \(\rightarrow\) Pi3DET (Drone)**
- **nuScenes (Vehicle) \(\rightarrow\) Pi3DET (Drone)**
- **Pi3DET (Vehicle) \(\rightarrow\) Pi3DET (Quadruped)**
- **nuScenes (Vehicle) \(\rightarrow\) Pi3DET (Quadruped)**
- **Pi3DET (Quadruped) \(\rightarrow\) Pi3DET (Drone)**
- **Pi3DET (Drone) \(\rightarrow\) Pi3DET (Quadruped)**

The two cross-dataset tasks are **nuScenes \(\rightarrow\) Pi3DET (Vehicle)** and **nuScenes \(\rightarrow\) KITTI**.

## 6. Pi3DET-Net: geometry alignment, feature alignment, and benchmark outcomes

To address the cross-platform problem posed by the Pi3DET dataset, the cited work proposes **Pi3DET-Net**, a two-stage framework consisting of **Pre-Adaptation (PA)** and **Knowledge-Adaptation (KA)** [2507.17665]. Its defining claim is that perspective-invariant 3D detection requires alignment at both **geometric** and **feature** levels.

The first geometry-level component is **Random Platform Jitter (RPJ)**. During pre-adaptation on the source platform, the method simulates roll and pitch instability by applying random rotations around the \(x\) and \(y\) axes. For a source point cloud \(\mathcal{P}^{\mathcal{S}}\) and box center \(\mathbf{c}\),
$$
\bar{\mathbf{p}} = \mathbf{R}(\Delta \phi, \Delta \theta)\,\mathbf{p}, \quad \bar{\mathbf{c}} = \mathbf{R}(\Delta \phi, \Delta \theta)\,\mathbf{c}.
$$
The appendix specifies default jitter uniformly sampled from **\([-5^\circ, +5^\circ]\)**. The second geometry-level component is **Virtual Platform Pose (VPP)**, which defines a virtual source-like pose on the target platform by setting roll and pitch to zero, preserving yaw, and aligning height to a vehicle-like coordinate frame:
$$
\bar{\mathbf{P}} = \bar{\mathbf{T}}\,{\mathbf{T}^{-1}}\,{\mathbf{P}}, \quad \bar{\mathbf{C}} = \bar{\mathbf{T}}\,{\mathbf{T}^{-1}}\,{\mathbf{C}}.
$$

Feature-level alignment is implemented through **Geometry-Aware Transformation Descriptor (GTD)** and **KL Probabilistic Feature Alignment (PFA)**. GTD predicts the synthetic RPJ parameters from source features, with rotation loss
$$
\mathcal{L}_{\mathrm{rot}} = \|\Delta\hat{\phi} - \Delta \phi\|^2 + \|\Delta\hat{\theta} - \Delta \theta\|^2.
$$
For region-of-interest features, PFA maps source and target RoI features into Gaussian latent distributions and aligns them with a KL objective:
$$
\mathcal{L}_\mathrm{KL} \;=\; D_\mathrm{KL}\Bigl[p(\mathbf{\xi}^{\mathcal{S}} \mid \mathbf{F}_r^{\mathcal{S}}) \,\Big\|\, p(\mathbf{\xi}^{\mathcal{T}} \mid \mathbf{F}_r^{\mathcal{T}}) \Bigr].
$$
The source RoI latent space is additionally supervised by
$$
\mathcal{L}_\mathrm{RoI} \;=\; \mathbb{E}_{\mathbf{\xi}^{\mathcal{S}}\sim p(\mathbf{\xi}^{\mathcal{S}}\mid F_r^{\mathcal{S}})} \bigl[-\log q(\mathbf{g}^{\mathcal{S}} \mid \mathbf{\xi}^{\mathcal{S}})\bigr].
$$

The overall PA objective is
$$
\mathcal{L}_\mathrm{PA} = \mathcal{L}_\mathrm{det} + \lambda_\mathrm{rot}\mathcal{L}_\mathrm{rot} + \lambda_\mathrm{RoI}\mathcal{L}_\mathrm{RoI},
$$
with \(\lambda_\mathrm{rot}=0.1\) and \(\lambda_\mathrm{RoI}=0.2\). During KA, the source branch is optimized with detection and RoI classification, the target branch with detection and KL alignment using \(\lambda_\mathrm{KL}=10^{-4}\), and the full objective is
$$
\mathcal{L}_\mathrm{KA} = \mathcal{L}^{\mathcal{T}}_\mathrm{KA}+ \mathcal{L}^{\mathcal{S}}_\mathrm{KA}.
$$

The benchmark evaluates a broad spectrum of detectors, including **PointPillar**, **SECOND-IOU**, **CenterPoint**, **PillarNet**, **Part-A\***, **TransFusion-L**, **HEDNet**, **SAFNet**, **PointRCNN**, **3DSSD**, **IA-SSD**, **DBQ-SSD**, **PV-RCNN**, **PV-RCNN++**, and **Voxel-RCNN**, together with adaptation baselines **ST3D**, **ST3D++**, **REDB**, and **MS3D++**. The paper states that on **Vehicle \(\rightarrow\) Drone** and **Vehicle \(\rightarrow\) Quadruped**, Pi3DET-Net improves mAP by **+11.84%** and **+12.03%** in one set of reported metrics. For **Vehicle \(\rightarrow\) Drone** using **Voxel-RCNN**, one benchmark table reports **65.63 / 44.62** at **AP@0.7**, and one reported setting gives **72.05 / 63.83** at **AP@0.5**. The paper further states that the method can outperform **target-only** training in some settings, plausibly because the target data are relatively small.

The cited ablations conclude that **RPJ**, **VPP**, **PFA**, and **GTD** each help, with the strongest results obtained by combining all four components. An RPJ angle study shows that **\(\pm 5^\circ\)** works best for **vehicle \(\rightarrow\) quadruped**, **\(\pm 3^\circ\)** works best for **vehicle \(\rightarrow\) drone**, and **\(\pm 8^\circ\)** hurts performance partly because points fall outside the detection range. The benchmark over **18 detectors** also yields broader observations: **hybrid grid-point detectors** often perform best on vehicles but degrade more strongly on quadruped and drone data, while **point-based detectors** are more balanced across platforms and often more viewpoint-robust.

## 7. Limitations, misconceptions, and research significance

A common misconception is that Pi3DET names a single mature technology. The cited literature does not support that reading. In detector physics, Pi3DET denotes a development program whose reported study is explicitly **not yet** the final 3D detector; it is a planar validation step for one of two proposed junction-forming strategies [2109.15010]. In machine perception, Pi3DET names a dataset and benchmark, while the accompanying **Pi3DET-Net** is explicitly stated to be **not truly viewpoint invariant** [2507.17665].

The detector-program literature identifies concrete constraints that remain unresolved: **0 V sensitivity** is still substantially lower than crystalline silicon detectors, low-temperature measurements become unstable below about **10 °C**, performance depends on **a-Si:H thickness**, and **TiO\(_2\)** underperforms **AZO** in leakage and transport quality [2109.15010]. The perception literature identifies a different set of limits: performance depends on the **strength of the base detector**, **predefined sensing ranges**, **pseudo-label quality**, and the **severity of viewpoint distortion**; failure cases include **misses under extreme viewpoint shifts**, **poor performance at long range**, and degradation when large pitch changes reduce the effective observable range [2507.17665].

Despite those limitations, both usages of Pi3DET occupy structurally analogous positions in their respective fields. In the detector case, the program reduces risk for a future 3D architecture by establishing that **charge-selective contacts** can function in **a-Si:H** detector structures, with **AZO** emerging as the preferred electron-selective contact [2109.15010]. In the perception case, the dataset and benchmark make **platform-level distribution shift** measurable and provide a concrete adaptation framework that aligns **geometry** and **feature distributions** across heterogeneous robotic platforms [2507.17665]. This suggests that, across both literatures, Pi3DET functions less as the name of a finished object than as a label for research programs organized around difficult 3D transfer problems: from planar to 3D detector fabrication in one case, and from vehicle-centric to genuinely multi-platform 3D perception in the other.

Source: https://www.emergentmind.com/topics/pi3det