Skip to content

Commit cb9f725

Browse files
author
Greg Cohen
committed
Added 10 new datastes
1 parent 87c2bbd commit cb9f725

File tree

10 files changed

+2262
-0
lines changed

10 files changed

+2262
-0
lines changed

datasets/2NMNIST.md

Lines changed: 173 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,173 @@
1+
---
2+
{
3+
"name": "2NMNIST",
4+
"aliases": [],
5+
"year": 2023,
6+
"modalities": [
7+
"Vision"
8+
],
9+
"sensors": [
10+
"ATIS"
11+
],
12+
"other_sensors": [],
13+
"category": "Object Detection, Classification, and Tracking",
14+
"tags": [
15+
"Classification",
16+
"Derived Datasets"
17+
],
18+
"description": "A Two-Label Classification Dataset based on the NMNIST",
19+
"dataset_properties": {
20+
"available_online": true,
21+
"has_real_data": true,
22+
"has_simulated_data": false,
23+
"has_ground_truth": false,
24+
"has_frames": true,
25+
"has_biases": false,
26+
"distribution_methods": [
27+
"Zenodo"
28+
],
29+
"file_formats": [
30+
"Binary"
31+
],
32+
"availability_comment": "",
33+
"dataset_links": [
34+
{
35+
"name": "Zenodo",
36+
"url": "https://zenodo.org/records/7847750",
37+
"format": "Binary",
38+
"doi": "10.5281/zenodo.7847750",
39+
"available": true
40+
}
41+
],
42+
"size_gb": 1.2,
43+
"size_type": "Compressed"
44+
},
45+
"paper": {
46+
"title": "Spikemoid: Updated Spike-based Loss Methods for Classification",
47+
"doi": "10.1109/IJCNN54540.2023.10191787",
48+
"authors": [
49+
"Michael Jurado",
50+
"Audrey Dunn",
51+
"Samuel Shapero"
52+
],
53+
"abstract": "Spiking Neural Networks (SNNs) have gained research attention in recent years due to their potential as low-power computing architectures for deployment on neuromorphic hardware. The introduction of offline training capabilities like Spike Layer Error Reassignment in Time (SLAYER) and advancements in the probabilistic interpretations of SNN output reinforce SNNs as a viable alternative to Artificial Neural Networks (ANNs). Spikemax was previously introduced as a family of differentiable loss methods which use windowed spike counts to form classification probabilities. We modify the spikemaxS loss method to use rates and a scaling parameter instead of counts to form scaled-spikemax. Our mathematical analysis shows that an appropriate scaling term can yield less coarse probability outputs from the SNN and help smooth the gradient of the loss during training. Experimentally, we show that scaled-spikemax achieves faster training convergence than spikemax and results in relative improvements of 4.2% and 9.9% in accuracy for NMNIST and N-TIDIGITS18, respectively We then extend scaled-spikemax to construct a spike-based loss function for multi-label classification called spikemoid. The viability of spikemoid is shown via the first known multi-label classification results on N-TIDIGITS18 and 2NMNIST, a novel variation of NMNIST that superimposes event-driven sensory data.",
54+
"open_access": false
55+
},
56+
"citation_counts": [
57+
{
58+
"source": "crossref",
59+
"count": 1,
60+
"updated": "2025-09-05T07:58:05.777621"
61+
},
62+
{
63+
"source": "scholar",
64+
"count": 2,
65+
"updated": "2025-09-05T07:58:07.296766"
66+
}
67+
],
68+
"links": [
69+
{
70+
"type": "github_page",
71+
"url": "https://github.com/audreydunn/spikemoid"
72+
},
73+
{
74+
"type": "paper",
75+
"url": "https://ieeexplore.ieee.org/document/10191787"
76+
}
77+
],
78+
"full_name": "",
79+
"additional_metadata": {},
80+
"referenced_papers": [
81+
{
82+
"doi": "10.1016/j.asoc.2018.05.012",
83+
"source": "crossref"
84+
},
85+
{
86+
"doi": "10.1103/PhysRevE.51.738",
87+
"source": "crossref"
88+
},
89+
{
90+
"doi": "10.1109/MSP.2019.2931595",
91+
"source": "crossref"
92+
},
93+
{
94+
"doi": "10.3389/fnins.2018.00023",
95+
"source": "crossref"
96+
},
97+
{
98+
"doi": "10.1109/SiPS52927.2021.00053",
99+
"source": "crossref"
100+
},
101+
{
102+
"doi": "10.1109/IJCNN55064.2022.9892379",
103+
"source": "crossref"
104+
},
105+
{
106+
"doi": "10.1016/S0361-9230(99)00161-6",
107+
"source": "crossref"
108+
},
109+
{
110+
"doi": "10.1109/CVPR.2017.781",
111+
"source": "crossref"
112+
},
113+
{
114+
"doi": "10.3389/fnins.2015.00437",
115+
"source": "crossref"
116+
},
117+
{
118+
"doi": "10.1109/JPROC.2021.3067593",
119+
"source": "crossref"
120+
},
121+
{
122+
"title": "Categorical reparameterization with gumbel-softmax",
123+
"source": "crossref"
124+
},
125+
{
126+
"title": "On calibration of modern neural networks",
127+
"source": "crossref"
128+
},
129+
{
130+
"title": "Training stochastic model recognition algorithms as networks can lead to maximum mutual information estimation of parameters",
131+
"source": "crossref"
132+
},
133+
{
134+
"title": "SLAYER: Spike layer error reassignment in time",
135+
"source": "crossref"
136+
},
137+
{
138+
"title": "1Asynchronous Binaural Spatial Audition Sensor With $2\\times 64\\times 4$ Channel Output",
139+
"source": "crossref"
140+
},
141+
{
142+
"title": "Spike-train level backpropagation fortraining deep recurrent spiking neural networks",
143+
"source": "crossref"
144+
},
145+
{
146+
"title": "Lapicque's introduction of the integrate-and-fire model neuron (1907)",
147+
"source": "crossref"
148+
}
149+
],
150+
"bibtex": {
151+
"pages": "1\u20137",
152+
"month": "jun",
153+
"year": 2023,
154+
"author": "Jurado, Michael and Dunn, Audrey and Shapero, Samuel",
155+
"publisher": "IEEE",
156+
"booktitle": "2023 International Joint Conference on Neural Networks (IJCNN)",
157+
"doi": "10.1109/ijcnn54540.2023.10191787",
158+
"url": "http://dx.doi.org/10.1109/IJCNN54540.2023.10191787",
159+
"title": "Spikemoid: Updated Spike-based Loss Methods for Classification",
160+
"type": "inproceedings",
161+
"key": "Jurado_2023"
162+
}
163+
}
164+
---
165+
166+
# Dataset Description
167+
168+
2NMNIST was constructed by randomly choosing distinct digits to be superimposed on each other. The superimposition is performed by time reversing one of the two digits andFig. 6. Example of a single frame of 2NMNIST imagery createtd from
169+
NMNIST frame superpositionoverlaying the two spike maps every timestep. Overlapping spike events are not summed in the resultant image.
170+
171+
The 2NMNIST dataset is devised to evaluate multilabel classification algorithms’ capability in detecting target classes that exhibit continuous mutual interference. In the training set there are 18730 superimposed samples and 22540 images with a single digit. In the testing set there are 3114 superimposed samples and 3772 with a single digit.
172+
173+
Note: The download link contains the combinations of digits used in generating the results for the paper. Code to reproduce the dataset is provided in the Github repository, however, regenerating the dataset on a different machine will yield different mixtures of digits.
Lines changed: 218 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,218 @@
1+
---
2+
{
3+
"name": "Aircraft Marshaling Signals Dataset",
4+
"aliases": [],
5+
"year": 2023,
6+
"modalities": [
7+
"Vision"
8+
],
9+
"sensors": [
10+
"DAVIS346"
11+
],
12+
"other_sensors": [
13+
"8 GHz UWB FMCW SISO radar senso"
14+
],
15+
"category": "Object Detection, Classification, and Tracking",
16+
"tags": [
17+
"Gesture Recognition",
18+
"Sensor Fusion",
19+
"Aircraft Marshalling"
20+
],
21+
"description": "Aircraft Marshaling Signals Dataset",
22+
"dataset_properties": {
23+
"available_online": true,
24+
"has_real_data": true,
25+
"has_simulated_data": false,
26+
"has_ground_truth": true,
27+
"has_frames": true,
28+
"has_biases": false,
29+
"distribution_methods": [
30+
"Zenodo"
31+
],
32+
"file_formats": [
33+
"Custom"
34+
],
35+
"availability_comment": "",
36+
"dataset_links": [
37+
{
38+
"name": "Zenodo",
39+
"url": "https://zenodo.org/records/10359770",
40+
"format": "Custom",
41+
"doi": "10.5281/zenodo.10359770",
42+
"available": true
43+
}
44+
],
45+
"size_gb": 5.9,
46+
"size_type": "Compressed"
47+
},
48+
"paper": {
49+
"title": "Aircraft Marshaling Signals Dataset of FMCW Radar and Event-Based Camera for Sensor Fusion",
50+
"doi": "10.1109/RadarConf2351548.2023.10149465",
51+
"authors": [
52+
"Leon M\u00fcller",
53+
"Manolis Sifalakis",
54+
"Sherif Eissa",
55+
"Amirreza Yousefzadeh",
56+
"Paul Detterer",
57+
"Sander Stuijk",
58+
"Federico Corradi"
59+
],
60+
"abstract": "The advent of neural networks capable of learning salient features from radar data has expanded the breadth of radar applications, often as an alternative sensor or a complementary modality to camera vision. Gesture recognition for command control is the most commonly explored application. Nevertheless, more suitable benchmarking datasets are needed to assess and compare the merits of the different proposed solutions. Furthermore, most current publicly available radar datasets used in gesture recognition provide little diversity, do not provide access to raw ADC data, and are not significantly challenging. To address these shortcomings, we created and made available a new dataset that combines two synchronized modalities: radar and dynamic vision camera of 10 aircraft marshaling signals at several distances and angles, recorded from 13 people. Moreover, we propose a sparse encoding of the time domain (ADC) signals that achieve a dramatic data rate reduction (>76%) while retaining the efficacy of the downstream FFT processing (<2% accuracy loss on recognition tasks). Finally, we demonstrate early sensor fusion results based on compressed radar data encoding in range-Doppler maps with dynamic vision data. This approach achieves higher accuracy than either modality alone.",
61+
"open_access": false
62+
},
63+
"citation_counts": [
64+
{
65+
"source": "crossref",
66+
"count": 5,
67+
"updated": "2025-09-05T15:17:11.111143"
68+
},
69+
{
70+
"source": "scholar",
71+
"count": 10,
72+
"updated": "2025-09-05T15:17:11.808209"
73+
}
74+
],
75+
"links": [
76+
{
77+
"type": "paper",
78+
"url": "https://ieeexplore.ieee.org/document/10149465"
79+
}
80+
],
81+
"full_name": "",
82+
"additional_metadata": {},
83+
"referenced_papers": [
84+
{
85+
"doi": "10.3389/fnbot.2021.688344",
86+
"source": "crossref"
87+
},
88+
{
89+
"doi": "10.1109/TGRS.2019.2929096",
90+
"source": "crossref"
91+
},
92+
{
93+
"doi": "10.1109/LSENS.2022.3173589",
94+
"source": "crossref"
95+
},
96+
{
97+
"doi": "10.23919/IRS54158.2022.9904979",
98+
"source": "crossref"
99+
},
100+
{
101+
"doi": "10.1561/9781680836523",
102+
"source": "crossref"
103+
},
104+
{
105+
"doi": "10.1109/MM.2018.112130359",
106+
"source": "crossref"
107+
},
108+
{
109+
"doi": "10.1109/LSSC.2019.2951423",
110+
"source": "crossref"
111+
},
112+
{
113+
"doi": "10.1109/JSSC.2014.2342715",
114+
"source": "crossref"
115+
},
116+
{
117+
"doi": "10.1049/el.2019.4153",
118+
"source": "crossref"
119+
},
120+
{
121+
"doi": "10.1145/2897824.2925953",
122+
"source": "crossref"
123+
},
124+
{
125+
"doi": "10.1109/TC.2022.3162708",
126+
"source": "crossref"
127+
},
128+
{
129+
"doi": "10.1109/TPAMI.2020.3008413",
130+
"source": "crossref"
131+
},
132+
{
133+
"doi": "10.1109/ICCV.2019.00058",
134+
"source": "crossref"
135+
},
136+
{
137+
"doi": "10.1109/CVPR.2017.781",
138+
"source": "crossref"
139+
},
140+
{
141+
"doi": "10.3390/rs14205177",
142+
"source": "crossref"
143+
},
144+
{
145+
"doi": "10.3389/fnins.2021.664208",
146+
"source": "crossref"
147+
},
148+
{
149+
"doi": "10.1109/LAWP.2022.3141869",
150+
"source": "crossref"
151+
},
152+
{
153+
"doi": "10.1109/EBCCSP53293.2021.9502357",
154+
"source": "crossref"
155+
},
156+
{
157+
"doi": "10.3390/electronics10121405",
158+
"source": "crossref"
159+
},
160+
{
161+
"doi": "10.3390/s21217298",
162+
"source": "crossref"
163+
},
164+
{
165+
"title": "Evolved neuromorphic radar-based altitude controller for an autonomous open-source blimp",
166+
"source": "crossref"
167+
},
168+
{
169+
"title": "Loihi: A neuromorphic manycore processor with on-chip learning",
170+
"source": "crossref"
171+
},
172+
{
173+
"title": "Improving the accuracy of spiking neural networks for radar gesture recognition through preprocessing",
174+
"source": "crossref"
175+
}
176+
],
177+
"bibtex": {
178+
"pages": "01\u201306",
179+
"month": "may",
180+
"year": 2023,
181+
"author": "M\u00fcller, Leon and Sifalakis, Manolis and Eissa, Sherif and Yousefzadeh, Amirreza and Detterer, Paul and Stuijk, Sander and Corradi, Federico",
182+
"publisher": "IEEE",
183+
"booktitle": "2023 IEEE Radar Conference (RadarConf23)",
184+
"doi": "10.1109/radarconf2351548.2023.10149465",
185+
"url": "http://dx.doi.org/10.1109/RadarConf2351548.2023.10149465",
186+
"title": "Aircraft Marshaling Signals Dataset of FMCW Radar and Event-Based Camera for Sensor Fusion",
187+
"type": "inproceedings",
188+
"key": "M_ller_2023"
189+
}
190+
}
191+
---
192+
193+
# Dataset Description
194+
195+
Synchronization of the two modalities
196+
197+
The PRI pulses from the radar have been hard-wired to the event stream of the DVS sensor, and timestamped using the DVS clock. Based on this signal the DVS event stream has been segmented such that groups of events (time-bins) of the DVS are mapped with individual radar pulses (chirps).
198+
199+
Data storage
200+
201+
DVS events (x,y coords and timestamps) are stored in structured arrays, and one such structured array object is associated with the data of a radar transmission (pulse/chirp). A radar transmission is a vector of 512 ADC levels that correspond to sampling points of chirping signal (FMCW radar) that lasts about ~1.3ms. Every 192 radar transmissions are stacked in a matrix called a radar frame (each transmission is a row in that matrix). A data capture (recording) consisting of some thousands of continuous radar transmissions is therefore segmented in a number of radar frames. Finally radar frames and the corresponding DVS structured arrays are stored in separate containers in a custom-made multi-container file format (extension .rad). We provide a (rad file) parser for extracting the data out of these files. There is one file per capture of continuous gesture recording of about 10s.
202+
203+
Note the number of 192 transmissions per radar frame is an ad-hoc segmentation that suits the purpose of obtaining sufficient signal resolution in a 2D FFT typical in radar signal processing, for the range resolution of the specific radar. It also served the purpose of fast streaming storing of the data during capture. For extracting individual data points for the dataset however, one can pool together (concat) all the radar frames from a single capture file and re-segment them according to liking. The data loader that we provide offers this, with a default of re-segmenting every 769 transmissions (about 1s of gesturing).
204+
205+
Data captures directory organization (radar8Ghz-DVS-marshaling_signals_20220901_publication_anonymized.7z)
206+
207+
The dataset captures (recordings) are organized in a common directory structure which encompasses additional metadata information about the captures.
208+
209+
dataset_dir/<stage>/<room>/<person>-<gesture>-<distance>/ofxRadar8Ghz_yyyy-mm-dd_HH-MM-SS.rad
210+
211+
Identifiers
212+
213+
stage [train, test].
214+
room: [conference_room, foyer, open_space].
215+
subject: [0-9]. Note that 0 stands for no person, and 1 for an unlabeled, random person (only present in test).
216+
gesture: ['none', 'emergency_stop', 'move_ahead', 'move_back_v1', 'move_back_v2', 'slow_down' 'start_engines', 'stop_engines', 'straight_ahead', 'turn_left', 'turn_right'].
217+
distance: ['xxx', '100', '150', '200', '250', '300', '350', '400', '450'] (in cm). Note that xxx is used for none gestures when there is no person present in front of the radar (i.e. background samples), or when a person is walking in front of the radar with varying distances but performing no gesture.
218+
The test data captures contain both subjects that appear in the train data as well as previously unseen subjects. Similarly the test data contain captures from the spaces that train data were recorded at, as well as from a new unseen open space.

0 commit comments

Comments
 (0)