POST
|
I have been able to train a deep learning model (Yolo v3) in ArcGIS Pro. The model seems fine, as can be seen below. It's not perfect but was trained with a small amount of data. However, it does not seem to work for inference. I am using the "Detect objects using deep learning" tool on the same image it was trained on to test the model, and specifying the model .emd. Everything runs fine and it produces a feature layer as expected. However, I've tried different combinations of parameters and the feature layer is always empty - no object detected. Even if I do the inference on a display extent where there a lot of objects (and this is the same imagery that was used for training so I should detect something). I am puzzled. Why doesn't my trained model detect anything? #### copy-pasted from .emd file { "Framework": "arcgis.learn.models._inferencing", "InferenceFunction": "[Functions]System\\DeepLearning\\ArcGISLearn\\ArcGISObjectDetector.py", "ModelConfiguration": "_yolov3_inference", "ModelType": "ObjectDetection", "ExtractBands": [ 0, 1, 2, 3 ], "ModelParameters": { "anchors": [ [ 8, 10 ], [ 24, 39 ], [ 73, 42 ], [ 71, 118 ], [ 180, 103 ], [ 128, 226 ], [ 276, 179 ], [ 235, 303 ], [ 368, 371 ] ], "n_bands": 4, "backbone": "DarkNet53", "backend": "pytorch" }, "Classes": [ { "Value": 1, "Name": "Pine", "Color": [ 124, 120, 159 ] } ], "ModelFormat": "NCHW", "MinCellSize": { "x": 0.029999999999998857, "y": 0.03000000000000922, "spatialReference": { "wkt": "PROJCS[\"WGS_1984_UTM_Zone_18N\",GEOGCS[\"GCS_WGS_1984\",DATUM[\"D_WGS_1984\",SPHEROID[\"WGS_1984\",6378137.0,298.257223563]],PRIMEM[\"Greenwich\",0.0],UNIT[\"Degree\",0.0174532925199433]],PROJECTION[\"Transverse_Mercator\"],PARAMETER[\"False_Easting\",500000.0],PARAMETER[\"False_Northing\",0.0],PARAMETER[\"Central_Meridian\",-75.0],PARAMETER[\"Scale_Factor\",0.9996],PARAMETER[\"Latitude_Of_Origin\",0.0],UNIT[\"Meter\",1.0]],VERTCS[\"unknown\",VDATUM[\"unknown\"],PARAMETER[\"Vertical_Shift\",0.0],PARAMETER[\"Direction\",1.0],UNIT[\"Meter\",1.0]]" } }, "MaxCellSize": { "x": 0.029999999999998857, "y": 0.03000000000000922, "spatialReference": { "wkt": "PROJCS[\"WGS_1984_UTM_Zone_18N\",GEOGCS[\"GCS_WGS_1984\",DATUM[\"D_WGS_1984\",SPHEROID[\"WGS_1984\",6378137.0,298.257223563]],PRIMEM[\"Greenwich\",0.0],UNIT[\"Degree\",0.0174532925199433]],PROJECTION[\"Transverse_Mercator\"],PARAMETER[\"False_Easting\",500000.0],PARAMETER[\"False_Northing\",0.0],PARAMETER[\"Central_Meridian\",-75.0],PARAMETER[\"Scale_Factor\",0.9996],PARAMETER[\"Latitude_Of_Origin\",0.0],UNIT[\"Meter\",1.0]],VERTCS[\"unknown\",VDATUM[\"unknown\"],PARAMETER[\"Vertical_Shift\",0.0],PARAMETER[\"Direction\",1.0],UNIT[\"Meter\",1.0]]" } }, "SupportsVariableTileSize": false, "ArcGISLearnVersion": "2.0.1", "monitored_valid_loss": 153.23129272460938, "ModelFile": "detect_pines_3cm_512.pth", "ImageHeight": 512, "ImageWidth": 512, "ImageSpaceUsed": "MAP_SPACE", "LearningRate": "slice('3.9811e-05', '3.9811e-04', None)", "ModelName": "YOLOv3", "backend": "pytorch", "accuracy": { "Pine": 0.4481547301495272 }, "resize_to": null, "IsMultispectral": true, "Bands": [ "", "", "", "" ], "ImageryType": "MultiSpectral", "NormalizationStats": { "band_min_values": [ 0.0, 0.0, 0.0, 0.0 ], "band_max_values": [ 255.0, 255.0, 255.0, 255.0 ], "band_mean_values": [ 94.2491226196289, 98.7134017944336, 65.65867614746094, 254.9991455078125 ], "band_std_values": [ 36.213584899902344, 37.29859161376953, 34.38850021362305, 0.46285802125930786 ], "scaled_min_values": [ 0.0, 0.0, 0.0, 0.0 ], "scaled_max_values": [ 1.0, 1.0, 1.0, 1.0 ], "scaled_mean_values": [ 0.3696043789386749, 0.3871113657951355, 0.2574850022792816, 0.9999966025352478 ], "scaled_std_values": [ 0.1420140564441681, 0.14626897871494293, 0.13485684990882874, 0.0018151294207200408 ] }, "DoNormalize": true }
... View more
12-14-2022
09:20 AM
|
1
|
3
|
732
|
POST
|
For a deep learning project that requires large amounts of training samples we rearly need to work as a team, with multiple workers contributing to the same training sample collection (shared feature class, .shp) and sharing the same classification schema (.ecs file). Is there a way to seamlessly share this with sync options via ArcGIS Online? Suppose say that one person updates a class (change the .ecs file) and other workers needed to have direct access to the new class. Instead of sending copies of local files around it would be really nice to sync online. And all samples should be ideally shared via a shared online feature service too. Is this possible?
... View more
02-19-2020
07:12 PM
|
0
|
0
|
711
|
POST
|
To add to this: Other companies are currently trying to fill that exact niche, for example Labelbox: https://labelbox.com/docs/tools/tiled-imagery Their image segmentation drawing tools are excellent https://labelbox.com/product/image-segmentation It's different than Collector in that it creates raster masks directly, but the drawing is much better from the user's perspective. But they're not there yet. First, you can't use it offline. Tablets are not supported (though it works). Not really meant for field use (i.e. no GPS etc). Also, zoom level is limited to 22, which is not enough. Also a problem I think on Collector (?), but not ArcMap where you can zoom as much as you want. With better polygon drawing tools (or why not, even raster masks like Labelbox!), Collector would be the best tool for this kind of image segmentation / annotation work where annotation must happen in the field. Le 23 janv. 2020 à 14:31, Etienne Laliberté <etienne.laliberte@umontreal.ca<mailto:etienne.laliberte@umontreal.ca>> a écrit : Hi Nick, Thanks. We take UAV high-res RGB imagery of vegetation. Then, we load the orthomosaic as a basemap in an ArcGIS online map. This map is shared with field users, who are field biologists. Field biologists then go in the field to specific areas (say, 10 x 10 m areas) and draw polygons around all individual plants/patches from the basemap, and identify to each species each plant/patch. The entire area needs to be fully covered. These polygons, annotated to plant species, are then used as training data for deep learning algorithms (semantic segmentation). The goal is to generate plant species map for the entire UAV orthomosaic, using those regions fully annotated as training data. Currently the only possible way we have found to do this is to use ArcMap on tablets, using a stylus. To make this operational, we need to draw with a stylus the contours of each plant. Drawing individual vertices is not good enough. It takes too long to use the current polygon drawing tool in Collector. Ideally we'd want to use the Apple Pencil on iPads on Collector, but that's not possible. Also, we're not necessarily standing at the vertex of the polygon we want to draw. In fact, we're generally not standing directly where we want to draw it. In ArcMap the freehand polygon tool is very useful, as well as the autocomplete freehand, because we need to completely cover small areas (say, 10 x 10 m areas on the ground) with polygons representing all of the different classes. We don't want overlap between polygons, or no empty space in between, hence the need for the autocomplete option. In ArcMap we create feature classes for every plant species / class to use. The only annoying thing with the autocomplete freehand is you can't use autocomplete freehand for different feature classes, even if these features are stored in the same shapefile. We can use the Trace tool to make sure polygons in a different feature class share exact boundaries with polygons from another feature class, but it's not as fast and a bit clumsy. We'd like the option to use autocomplete polygon (freehand or not) on polygons from different feature classes. If we could all do that in Collector, it would be the perfect tool. With deep learning methods on the rise, and the need for such training data for semantic segmentation algorithms (e.g. convolutional neural networks), I can see how many plant ecologists / field biologists will be looking for tools to do exactly that. Collector is well-positioned to fill that niche, if it could only upgrade its polygon drawing capability. Cheers, Etienne Le 23 janv. 2020 à 12:31, Nick Black <geonet@esri.com<mailto:geonet@esri.com>> a écrit : GeoNet, The Esri Community | GIS and Geospatial Professional Community<https://community.esri.com/?et=watches.email.thread> Re: Collector: Anyone else find moving the map to add features awkward? reply from Nick Black<https://community.esri.com/people/NBlack-esristaff?et=watches.email.thread> in Collector for ArcGIS - View the full discussion<https://community.esri.com/message/903886-re-collector-anyone-else-find-moving-the-map-to-add-features-awkward?commentID=903886&et=watches.email.thread#comment-903886>
... View more
01-23-2020
11:38 AM
|
2
|
0
|
1398
|
POST
|
Hi Nick, Thanks. We take UAV high-res RGB imagery of vegetation. Then, we load the orthomosaic as a basemap in an ArcGIS online map. This map is shared with field users, who are field biologists. Field biologists then go in the field to specific areas (say, 10 x 10 m areas) and draw polygons around all individual plants/patches from the basemap, and identify to each species each plant/patch. The entire area needs to be fully covered. These polygons, annotated to plant species, are then used as training data for deep learning algorithms (semantic segmentation). The goal is to generate plant species map for the entire UAV orthomosaic, using those regions fully annotated as training data. Currently the only possible way we have found to do this is to use ArcMap on tablets, using a stylus. To make this operational, we need to draw with a stylus the contours of each plant. Drawing individual vertices is not good enough. It takes too long to use the current polygon drawing tool in Collector. Ideally we'd want to use the Apple Pencil on iPads on Collector, but that's not possible. Also, we're not necessarily standing at the vertex of the polygon we want to draw. In fact, we're generally not standing directly where we want to draw it. In ArcMap the freehand polygon tool is very useful, as well as the autocomplete freehand, because we need to completely cover small areas (say, 10 x 10 m areas on the ground) with polygons representing all of the different classes. We don't want overlap between polygons, or no empty space in between, hence the need for the autocomplete option. In ArcMap we create feature classes for every plant species / class to use. The only annoying thing with the autocomplete freehand is you can't use autocomplete freehand for different feature classes, even if these features are stored in the same shapefile. We can use the Trace tool to make sure polygons in a different feature class share exact boundaries with polygons from another feature class, but it's not as fast and a bit clumsy. We'd like the option to use autocomplete polygon (freehand or not) on polygons from different feature classes. If we could all do that in Collector, it would be the perfect tool. With deep learning methods on the rise, and the need for such training data for semantic segmentation algorithms (e.g. convolutional neural networks), I can see how many plant ecologists / field biologists will be looking for tools to do exactly that. Collector is well-positioned to fill that niche, if it could only upgrade its polygon drawing capability. Cheers, Etienne Le 23 janv. 2020 à 12:31, Nick Black <geonet@esri.com<mailto:geonet@esri.com>> a écrit : GeoNet, The Esri Community | GIS and Geospatial Professional Community<https://community.esri.com/?et=watches.email.thread> Re: Collector: Anyone else find moving the map to add features awkward? reply from Nick Black<https://community.esri.com/people/NBlack-esristaff?et=watches.email.thread> in Collector for ArcGIS - View the full discussion<https://community.esri.com/message/903886-re-collector-anyone-else-find-moving-the-map-to-add-features-awkward?commentID=903886&et=watches.email.thread#comment-903886>
... View more
01-23-2020
11:32 AM
|
3
|
0
|
1398
|
POST
|
I can run ArcMap 10.7 on a Surface Pro X (which has ARM64 chip, running Windows 10) while connected to our institution network. But I want to run it offline in the field on the tablet. ArcMap seems to run OK (not super smooth but usable) with the built-in emulator for 32-bit programs. However, ArcGIS Administrator does not work properly. I have an active single-use (Advanced) licence, and I can run the activation process all the way to the end (I even get the message the the activation in complete), but I click 'End', and I get back to the start, and the software is not activated. Therefore I cannot run ArcMap while offline. Questions (1) will you fix this issue and (2) will you soon release an ArcMap version compatible with ARM64? thanks.
... View more
01-11-2020
09:58 AM
|
1
|
1
|
1399
|
POST
|
Collector is made for field use. A tablet (eg iPads) with a stylus is a great device for field data collection. The ability to draw freehand polygons and autocomplete freehand polygons with a stylus directly on the basemap seems like a must-have for an app like this. It's disappointing that this is not possible.
... View more
01-08-2020
06:39 PM
|
0
|
0
|
309
|
POST
|
I totally agree - please consider adding freehand and autocomplete freehand polygons to Collector. Our workflow needs it. And we also digitise polygons without necessarily standing next to the point.
... View more
01-08-2020
06:32 PM
|
4
|
4
|
1398
|
Title | Kudos | Posted |
---|---|---|
1 | 12-14-2022 09:20 AM | |
2 | 01-23-2020 11:38 AM | |
3 | 01-23-2020 11:32 AM | |
1 | 01-11-2020 09:58 AM | |
4 | 01-08-2020 06:32 PM |
Online Status |
Offline
|
Date Last Visited |
12-15-2022
03:31 PM
|