MiliPoint Dataset: mmWave Radar Benchmark

• MiliPoint is a large-scale mmWave radar dataset that uses sparse 3D point clouds for privacy-preserving human activity recognition across identification, action classification, and keypoint estimation tasks.
• The dataset employs a fixed-size point cloud representation with zero-padding and temporal stacking to handle noise and sparsity, facilitating robust point-set deep learning.
• Benchmark evaluations with architectures like DGCNN, Pointformer, PointNet++, and PointMLP reveal high identification accuracy but consistently challenging action classification performance.

Searching arXiv for MiliPoint and directly related follow-up work. MiliPoint is a large-scale, point-cloud–based mmWave radar dataset for human activity sensing that was introduced to provide a standardized benchmark for point-set deep learning on RF sensing data. It is designed around three core tasks—person identification, action classification, and human keypoint estimation—and uses sparse 3D point clouds derived from millimetre-wave radar rather than camera imagery, making the sensing modality intrinsically non-visual and less intrusive (Cui et al., 2023). Subsequent work has also used MiliPoint as a testbed for privacy-aware human action recognition pipelines based on DBSCAN, the Hungarian algorithm, and Kalman filtering, emphasizing the dataset’s role in studying how classical point-cloud processing can stabilize sparse and noisy radar measurements for downstream deep models (Tunau et al., 14 Aug 2025).

1. Definition and scope

MiliPoint is a mmWave radar point-cloud dataset specifically designed for human activity recognition and related sensing tasks. In the original dataset paper, it is presented as a large-scale, open dataset for the community to explore how mmWave radars can be utilised for human activity recognition, with point clouds used as the canonical data representation rather than image-like radar products such as range–Doppler heatmaps (Cui et al., 2023).

The dataset targets three tasks within a unified benchmark. The first is identification, in which the objective is to infer subject identity from stacked radar point clouds. The second is action classification, in which a temporal sequence of radar point clouds is mapped to an action label. The third is keypoint estimation, in which the model regresses 3D human keypoints in radar coordinates (Cui et al., 2023). This joint coverage distinguishes MiliPoint from earlier mmWave datasets that predominantly targeted only keypoint or pose estimation, or only action classification.

A central characteristic of MiliPoint is that each radar frame is represented only by 3D coordinates $(x, y, z)$ of radar returns. The point clouds are sparse, noisy, and temporally unstable, reflecting the underlying sensing physics of mmWave RF reflections and the limitations of CFAR-based point extraction (Cui et al., 2023). This makes the dataset simultaneously a benchmark for privacy-preserving sensing and for robust point-set learning under severe data sparsity.

2. Sensor platform and acquisition protocol

MiliPoint uses a TI IWR1843 commercial mmWave radar operating in the 77–81 GHz band, with 3 transmit (TX) and 4 receive (RX) antennas in a time-division multiplexing MIMO configuration (Cui et al., 2023). The reported radar signal configuration includes a chirp duration of $T_{\text{chirp}} = 100\,\mu\text{s}$, a chirp slope of $S = 40\,\text{MHz}/\mu\text{s}$, a total bandwidth of $B = 4\,\text{GHz}$, an ADC sampling rate of $f_s = 5\,\text{MHz}$, and CFAR thresholds of 10 dB in both range and Doppler domains (Cui et al., 2023). The range resolution is reported as about 4 cm.

The acquisition setup also includes a Zed 2 stereo camera, used only to obtain ground-truth 3D human keypoints. The camera provides disparity and depth maps, and a pose-estimation network estimates 3D skeletons that are then transformed into the radar coordinate system via calibration (Cui et al., 2023). The released dataset contains only radar point clouds and derived keypoints, while camera images are discarded after keypoint generation.

Data were collected in an indoor environment, controlled and largely free of external RF interference. Participants performed a 30-minute low-intensity cardio workout consisting of 49 distinct actions, with each action lasting approximately 30 s and separated by 10 s breaks (Cui et al., 2023). Participants remained within a 1 m $\times$ 1 m square in front of the radar. The radar was placed 0.65 m from the area center at a height of 1.0 m, while the camera was placed 3.0 m from the area center at a height of 0.7 m. Radar point clouds and camera frames were both captured at 24 fps and synchronized by timestamp (Cui et al., 2023).

This acquisition protocol establishes a controlled benchmark rather than a highly unconstrained naturalistic corpus. The dataset therefore emphasizes reproducibility and standardized evaluation over broad viewpoint or environmental variation.

3. Data representation, scale, and annotations

MiliPoint contains 545,059 radar frames, or 545,000 frames as summarized in later follow-up work, recorded from 11 participants and covering 49 distinct human actions (Cui et al., 2023, Tunau et al., 14 Aug 2025). The participant cohort consists of 4 female and 7 male subjects, with average height $171.84 \pm 10.41$ cm and average weight $67.73 \pm 13.08$ kg; all participants had no mobility impairments (Cui et al., 2023).

Each radar frame is represented as a point cloud

$d \in \mathbb{R}^{N \times 3},$

where each row is a Cartesian coordinate $(x, y, z)$ corresponding to a detected reflection in metric 3D space (Cui et al., 2023). Only the coordinate triplet is retained in the released representation; intensity, Doppler, or related radar attributes are not explicitly included in the dataset representation described in the paper.

Because the number of points $T_{\text{chirp}} = 100\,\mu\text{s}$0 varies from frame to frame, the dataset uses a fixed-size representation. If a frame contains more than a maximum number of points $T_{\text{chirp}} = 100\,\mu\text{s}$1, it is randomly downsampled; if it contains fewer than $T_{\text{chirp}} = 100\,\mu\text{s}$2, it is zero-padded. Each frame is then represented as

$T_{\text{chirp}} = 100\,\mu\text{s}$3

To encode motion, consecutive frames are stacked, yielding an input tensor

$T_{\text{chirp}} = 100\,\mu\text{s}$4

where $T_{\text{chirp}} = 100\,\mu\text{s}$5 is the stacking length (Cui et al., 2023).

The dataset provides labels for three tasks:

Task	Label space	Notes
Identification	$T_{\text{chirp}} = 100\,\mu\text{s}$6	11 participant identities
Action classification	$T_{\text{chirp}} = 100\,\mu\text{s}$7	49 actions recorded; 39 used for classification
Keypoint estimation	$T_{\text{chirp}} = 100\,\mu\text{s}$8 or $T_{\text{chirp}} = 100\,\mu\text{s}$9	Two keypoint granularities

For keypoint estimation, MiliPoint defines both a 9-point and an 18-point setting. The 9-point setting includes Right Shoulder, Right Elbow, Left Shoulder, Left Elbow, Right Hip, Right Knee, Left Hip, Left Knee, and Head. The 18-point setting includes finer landmarks such as Nose, Neck, wrists, ankles, eyes, and ears, with Head excluded due to the finer facial landmark decomposition (Cui et al., 2023).

Recommended stacking hyperparameters are task-dependent: $S = 40\,\text{MHz}/\mu\text{s}$0 for identification, $S = 40\,\text{MHz}/\mu\text{s}$1 for both keypoint estimation tasks, and $S = 40\,\text{MHz}/\mu\text{s}$2 for action classification (Cui et al., 2023). Follow-up work on preprocessing additionally states that temporal stacking is used so that each frame contains up to 1100 points, following Cui et al. (Tunau et al., 14 Aug 2025).

4. Benchmark tasks and baseline performance

The original benchmark uses an 80% training, 10% validation, and 10% test split over samples, with evaluation performed separately for the three task families (Cui et al., 2023). The principal metrics are Top-1 accuracy for identification, Top-1 and Top-3 accuracy for action classification, and Mean Localization Error (MLE) in centimeters for keypoint estimation: $S = 40\,\text{MHz}/\mu\text{s}$3 Here, $S = 40\,\text{MHz}/\mu\text{s}$4 and $S = 40\,\text{MHz}/\mu\text{s}$5 denote predicted and ground-truth keypoint coordinates (Cui et al., 2023).

The baseline benchmark evaluates four representative point-based architectures: DGCNN, Pointformer, PointNet++, and PointMLP (Cui et al., 2023). Their reported results are as follows.

Model	Identification Top1 (%)	Action Top1 (%)	Action Top3 (%)
DGCNN	$S = 40\,\text{MHz}/\mu\text{s}$6	$S = 40\,\text{MHz}/\mu\text{s}$7	$S = 40\,\text{MHz}/\mu\text{s}$8
Pointformer	$S = 40\,\text{MHz}/\mu\text{s}$9	$B = 4\,\text{GHz}$0	$B = 4\,\text{GHz}$1
PointNet++	$B = 4\,\text{GHz}$2	$B = 4\,\text{GHz}$3	$B = 4\,\text{GHz}$4
PointMLP	$B = 4\,\text{GHz}$5	$B = 4\,\text{GHz}$6	$B = 4\,\text{GHz}$7

For keypoint estimation, the corresponding MLE results are 16.53 ± 0.11 cm and 18.51 ± 0.03 cm for DGCNN, 14.99 ± 0.03 cm and 17.03 ± 0.13 cm for Pointformer, 13.55 ± 0.03 cm and 14.94 ± 0.03 cm for PointNet++, and 13.12 ± 0.30 cm and 14.11 ± 0.22 cm for PointMLP on the 9-point and 18-point settings respectively (Cui et al., 2023).

These results establish a characteristic task hierarchy. Identification is comparatively easy, with all models exceeding 75% Top-1 accuracy and PointMLP reaching approximately 96%. Keypoint estimation is also effective, with localization errors around 13–18 cm. Action classification is substantially harder, with the best Top-1 result at only about 34–35%, indicating that extracting action semantics from sparse mmWave point-cloud sequences remains far from solved (Cui et al., 2023).

This difficulty profile became especially prominent in later work, where training directly on raw MiliPoint point clouds yielded low action-recognition baselines—13.61% for DGCNN, 29.27% for Pointformer, 34.45% for PointNet++, and 18.37% for PointMLP—before explicit preprocessing was applied (Tunau et al., 14 Aug 2025).

5. Noise characteristics and preprocessing-oriented use

A defining property of MiliPoint is that the point clouds are sparse and noisy. The original dataset paper attributes this to thermal noise, CFAR thresholding, multipath reflections, clutter, and incomplete sampling of body surfaces, all of which produce variable and unstable point distributions across frames (Cui et al., 2023). The later privacy-aware HAR study operationalizes this sparsity more concretely by describing a temporally stacked representation in which each frame contains up to 1100 points, but many of these are zero-padding or non-informative detections (Tunau et al., 14 Aug 2025).

In that follow-up pipeline, each frame is divided into 5 contiguous segments of 220 points. Null-value removal is performed using the threshold

$B = 4\,\text{GHz}$8

which discards points near the origin treated as zero padding (Tunau et al., 14 Aug 2025). After this cleaning step, the average effective density becomes approximately 63 non-zero points per segment in the 5-segment configuration, 132 non-zero points per segment in a 2-segment configuration, and 32 non-zero points per segment in a 10-segment configuration.

The same study evaluates three classical processing methods on MiliPoint: DBSCAN, the Hungarian algorithm, and Kalman filtering (Tunau et al., 14 Aug 2025). For DBSCAN, the selected parameters are $B = 4\,\text{GHz}$9 and $f_s = 5\,\text{MHz}$0, with a vertical weighting factor $f_s = 5\,\text{MHz}$1 that de-emphasizes the $f_s = 5\,\text{MHz}$2-axis in the distance metric. Cluster centroids are computed as

$f_s = 5\,\text{MHz}$3

and the Hungarian algorithm then solves a centroid-assignment problem across segments using the Euclidean cost matrix

$f_s = 5\,\text{MHz}$4

Kalman filtering is applied in the horizontal plane with state vector

$f_s = 5\,\text{MHz}$5

and observation vector

$f_s = 5\,\text{MHz}$6

with parameters tuned by Bayesian Optimization to $f_s = 5\,\text{MHz}$7, $f_s = 5\,\text{MHz}$8, and $f_s = 5\,\text{MHz}$9, alongside a 2.0 meter association threshold (Tunau et al., 14 Aug 2025). This pipeline reduces each sample from 1100 raw points to approximately 60–300 points in the final human cluster.

The study reports a representative four-track example in which Track 0 has RMSE = 0.9317 and median distance = 0.0964 m, Track 1 has RMSE = 1.3630 and median distance = 1.4454 m, Track 2 has RMSE = 1.0531 and median distance = 1.0316 m, and Track 3 has RMSE = 0.7158 and median distance = 0.4745 m; merging Tracks 0 and 3 yields the cluster judged both accurate and spatially aligned with the human ground truth (Tunau et al., 14 Aug 2025).

This preprocessing-focused literature shows that MiliPoint functions not only as a benchmark for end-to-end point-set networks, but also as a benchmark for denoising, clustering, association, and tracking strategies tailored to sparse radar point clouds.

6. Comparison with earlier mmWave datasets and distinctive contributions

The original paper explicitly positions MiliPoint against prior mmWave datasets such as mmPose, MARS, HuPR, mRI, CubeLearn, and RadHAR (Cui et al., 2023). In that comparison, the principal axes are task coverage, number of participants, dataset size, and number of actions.

Dataset	Tasks	Participants	Size (frames)	# Actions
mmPose	K	2	15k	4
MARS	K	4	40k	10
HuPR	K	6	141k	3
mRI	K	20	160k	12
CubeLearn	A	8	1k	6
RadHAR	A	2	167k	5
MiliPoint	A, I, K	11	545k	49

Within this comparison, MiliPoint is described as the largest mmWave HAR dataset to date, with 545k radar frames in total, 49 distinct actions, and support for all three key HAR tasks in one dataset (Cui et al., 2023). The paper further states that MiliPoint is 4.08× more actions than the most diverse earlier dataset and about 3.26× larger than the largest prior datasets in frame count.

The significance of this comparison is methodological rather than merely statistical. Earlier datasets generally focused on a single task, often pose estimation only or action classification only, and many used non-point-cloud representations. MiliPoint instead standardizes a point-cloud representation and supplies a common platform for evaluating GNN-based, transformer-based, hierarchical set-abstraction, and MLP-based point architectures under the same sensing setup (Cui et al., 2023).

A plausible implication is that MiliPoint helped shift mmWave HAR benchmarking toward point-set learning as a first-class problem, rather than treating radar outputs primarily as 2D image-like tensors. The follow-up preprocessing study reinforces this reading by showing that classical geometric processing can materially change downstream classification performance on the same dataset (Tunau et al., 14 Aug 2025).

7. Privacy, limitations, and access

MiliPoint is repeatedly framed as a privacy-preserving alternative to camera-based sensing because mmWave radar captures only RF reflections as sparse 3D points and does not encode facial appearance, clothing, or other direct visual identifiers (Cui et al., 2023, Tunau et al., 14 Aug 2025). The dataset paper states that participants were fully informed and signed consent forms, that camera images were used only to compute keypoints and discarded immediately thereafter, that the released data contain only radar point clouds and derived keypoints, and that the study was approved by the University of Bristol Faculty of Engineering Research Ethics Committee (reference 12802) (Cui et al., 2023).

The dataset nevertheless has several explicit limitations. It is based on a single-radar indoor setup in a low-clutter environment; participants generally face the radar; angular variation is limited; movements are largely low-intensity cardio emphasizing limb motion; and more complex postures such as sitting or lying are not represented (Cui et al., 2023). The radar’s 3 TX / 4 RX arrangement is optimized for automotive settings and provides stronger azimuth than elevation resolution, which may reduce sensitivity to actions involving critical height changes. The paper also notes that multipath effects are not fully addressed, and that experiments are primarily line-of-sight (Cui et al., 2023).

Later work adds a further practical limitation: strong performance on MiliPoint action recognition may depend on explicit preprocessing. On the reported benchmark, Kalman filtering alone yields 98.45%, 96.11%, 98.48%, and 98.57% accuracy across DGCNN, Pointformer, PointNet++, and PointMLP respectively; DBSCAN alone yields 98.19%, 96.14%, 98.54%, and 98.42%; Hungarian + DBSCAN remains similarly strong; while the full DS + KM + HG pipeline reaches 95.26%, 91.45%, 95.67%, and 95.31% but with a much higher preprocessing time of 0.2369 s per sample (Tunau et al., 14 Aug 2025). This suggests that benchmark outcomes on MiliPoint can be highly sensitive to the front-end processing regime.

MiliPoint is available through the GitHub repository https://github.com/yizzfz/MiliPoint/, with raw data hosted on Google Drive at https://drive.google.com/file/d/1rq8yyokrNhAGQryx7trpUqKenDnTI6Ky/ (Cui et al., 2023). The dataset is released under the MIT license. The original repository includes dataset documentation, code for loading and preprocessing point clouds, baseline models, training scripts, and pre-trained weights (Cui et al., 2023). The later preprocessing pipeline is also publicly released at https://github.com/Maimunatunau/Human-Action-Recognition-HAR-using-mmWave-Radar (Tunau et al., 14 Aug 2025).

Taken together, these properties make MiliPoint a reference dataset for privacy-aware mmWave HAR research: large enough to support comparative benchmarking, technically standardized enough to support reproducibility, and challenging enough to expose open problems in action recognition, temporal modeling, denoising, and robust point-cloud learning (Cui et al., 2023).