A Light-Weight Neural Network for Fast and Interactive Edge Detection in 3D Point Clouds

ABC Dataset (noisy version)

This experiment consists in classifying edge points on a noisy version of the chunk 0000 (7167 models) of the ABC Dataset. We randomly selected 200 and 50 models for training and validation, respectively:

Training models: 0037, 0039, 0044, 0048, 0101, 0116, 0124, 0137, 0143, 0150, 0156, 0189, 0202, 0230, 0233, 0254, 0299, 0310, 0324, 0342, 0374, 0382, 0387, 0392, 0435, 0457, 0467, 0480, 0501, 0502, 0513, 0533, 0542, 0550, 0551, 0553, 0584, 0603, 0618, 0640, 0646, 0654, 0662, 0664, 0673, 0683, 0686, 0688, 0746, 0753, 0794, 0840, 0871, 0901, 0915, 0925, 0950, 0959, 0980, 0982, 1001, 1022, 1035, 1043, 1054, 1058, 1080, 1095, 1139, 1142, 1143, 1154, 1173, 1193, 1197, 1203, 1226, 1231, 1242, 1250, 1263, 1299, 1301, 1314, 1341, 1361, 1369, 1374, 1387, 1390, 1393, 1401, 1403, 1405, 1413, 1418, 1452, 1473, 1521, 1532, 1536, 1538, 1559, 1589, 1599, 1607, 1608, 1611, 1613, 1617, 1627, 1634, 1675, 1697, 1700, 1713, 1722, 1724, 1729, 1739, 1744, 1751, 1761, 1811, 1840, 1841, 1848, 1877, 1896, 1898, 1903, 1913, 1944, 1950, 1954, 1955, 1961, 1967, 2034, 2036, 2046, 2082, 2089, 2107, 2118, 2131, 2135, 2138, 2161, 2170, 2177, 2189, 2201, 2207, 2236, 2245, 2295, 2304, 2310, 2329, 2340, 2354, 2358, 2389, 2425, 2426, 2427, 2430, 2432, 2435, 2438, 2452, 2472, 2487, 2492, 2499, 2517, 2520, 2538, 2540, 2584, 2586, 2596, 2609, 2678, 2685, 2686, 2690, 2714, 2719, 2726, 2732, 2744, 2763, 2768, 2771, 2776, 2777, 2799, 2800,
Validation models: 0008, 0091, 0138, 0198, 0277, 0353, 0402, 0487, 0541, 0576, 0643, 0667, 0713, 0844, 0939, 0994, 1051, 1106, 1155, 1217, 1258, 1323, 1384, 1402, 1433, 1534, 1594, 1612, 1672, 1716, 1743, 1813, 1884, 1937, 1957, 2038, 2117, 2139, 2192, 2248, 2337, 2392, 2431, 2465, 2507, 2572, 2667, 2703, 2738, 2775.

Noisy samples are obtained by applying displacement noise in the direction of the normal vector, with magniture=4% of the object bounding box

Examples of 3d models

Model description

Select a classification in the list:

Select a model in the list:

Legend

Flat
Sharp
Smooth

Interactive WebGL viewer

(click, drag and scroll to interact with the 3D model)

Statistical analysis

Indicators: median value per method

Click on a column header to change sorting.

Search:

Method	Precision	Recall	MCC	TP	FP	TN	FN	F1	Accuracy	Selectivity	IoU
PCED (trained on Default - P = Sharp+Smooth)	0.231	0.920	0.358	0.071	0.982	0.929	0.018	0.366	0.768	0.761	0.222
Edge-aware Point set Consolidation Network (Pre-trained)	0.247	0.687	0.317	0.052	0.895	0.948	0.105	0.358	0.824	0.837	0.216
PCED (trained on Default)	0.323	0.375	0.267	0.025	0.551	0.975	0.449	0.319	0.903	0.945	0.185
PCED-2C (trained on EC)	0.163	0.950	0.236	0.090	0.993	0.910	0.007	0.275	0.633	0.615	0.158
PCED-2C (trained on ABC)	0.121	0.793	0.181	0.082	0.967	0.918	0.033	0.213	0.634	0.626	0.117
PCPNet (trained on ABC)	0.159	0.363	0.155	0.025	0.726	0.975	0.274	0.204	0.833	0.878	0.113

Showing 1 to 6 of 6 entries

Indicators distributions per method

This plot shows, for each method and training set, the distributions of several indicators for the whole dataset. Use the drop-down menue below to change the indicator (both the description and the plot should be updated). Note that the plot is interactive: change zoom, range and access method full name using mouse interaction.
Current indicator:

Precision (also denoted positive predictive value -- PPV) measures the proportion of positive identifications that was actually correct (higher is better). It is defined as:

precision = \frac{T P}{T P + F P}

Bad (indicator 25% smallest values) and good (indicator 10% higher values) areas are highlighted in the plot in red and green, respectively. Efficient methods should have few samples in the red area (the lower, the better), while having most of its samples in the green area (the higher, the better).

Precision/Recall

This plot shows a representative score per approach: each method is shown as point sample, such that (select mean or median below):

Its position is the mean precision/recall score for the whole dataset,
Its position is the median precision/recall score for the whole dataset.

The plot is interactive: change zoom, range and access method full name using mouse interaction.

Visualization type:

Precision/Recall distributions - Methods comparisons

These graphs show the precision/recall values as scatter plot (each 3D model is a sample) and its associated density function for the whole dataset.

Left dataset:

Right dataset: