release commit
authorManu Mathew <a0393608@ti.com>
Thu, 9 Jan 2020 05:19:47 +0000 (10:49 +0530)
committerManu Mathew <a0393608@ti.com>
Thu, 9 Jan 2020 05:19:47 +0000 (10:49 +0530)
README.md
docs/Semantic_Segmentation.md
run_segmentation.sh

index dc9ef7aa39c0ff9ed46cf6acd992a7663c61d8ca..6d526efc215614ba1878eb9c8f7c3fd8c60393fa 100644 (file)
--- a/README.md
+++ b/README.md
@@ -42,7 +42,7 @@ The following examples are currently available. Click on each of the links below
 Some of the common training and validation commands are provided in shell scripts (.sh files) in the root folder.
 
 #### **Issue Tracker**: 
-- You can file issues or ask questions at **e2e**: [https://e2e.ti.com/support/j721e/f/1026/tags/jacinto_2D00_ai_2D00_devkit](https://e2e.ti.com/support/j721e/f/1026/tags/jacinto_2D00_ai_2D00_devkit).<br>
+- You can file issues or ask questions at **e2e**: [https://e2e.ti.com/support/processors/f/791/tags/jacinto_2D00_ai_2D00_devkit](https://e2e.ti.com/support/processors/f/791/tags/jacinto_2D00_ai_2D00_devkit).<br>
 - The part number should be filled in as **TDA4VM**. Also, kindly include **jacinto-ai-devkit** in the tags (at the end of the page as you create a new issue) so that we get notified quickly. 
 - If you do not get a reply within two days, you can contact us at: jacinto-ai-devkit@list.ti.com
 
index a59781351333b2e3c2934a438dde65f9b9f8ab78..1d4826f3a2c87f56cee5e11479efb7263c404a17 100644 (file)
@@ -79,28 +79,30 @@ Inference can be done as follows (fill in the path to the pretrained model):<br>
 
 ### Cityscapes Segmentation
 
-|Dataset    |Mode Architecture         |Backbone Model|Backbone Stride|Resolution |Complexity (GigaMACS)|MeanIoU%  |Model Configuration Name                    |
-|---------  |----------                |-----------   |-------------- |-----------|--------             |----------|----------------------------------------    |
-|Cityscapes |FPNPixel2Pixel with DWASPP|MobileNetV2   |64             |768x384    |0.99                 |62.43     |fpn_pixel2pixel_aspp_mobilenetv2_tv_fd    |
-|Cityscapes |DeepLabV3Lite with DWASPP |MobileNetV2   |16             |768x384    |**3.54**             |**69.13** |**deeplabv3lite_mobilenetv2_tv**            |
-|Cityscapes |FPNPixel2Pixel            |MobileNetV2   |32(\*2\*2)     |768x384    |3.66                 |70.30     |fpn_pixel2pixel_mobilenetv2_tv              |
-|Cityscapes |FPNPixel2Pixel with DWASPP|MobileNetV2   |32             |768x384    |3.84                 |70.39     |fpn_pixel2pixel_aspp_mobilenetv2_tv         |
-|Cityscapes |FPNPixel2Pixel            |MobileNetV2   |64(\*2\*2)     |1536x768   |3.85                 |69.82     |fpn_pixel2pixel_mobilenetv2_tv_fd         |
-|Cityscapes |FPNPixel2Pixel with DWASPP|MobileNetV2   |64             |1536x768   |**3.96**             |**71.28** |**fpn_pixel2pixel_aspp_mobilenetv2_tv_fd**|
-|Cityscapes |FPNPixel2Pixel with DWASPP|MobileNetV2   |64             |2048x1024  |7.03                 |72.67     |fpn_pixel2pixel_aspp_mobilenetv2_tv_fd    |
-|Cityscapes |DeepLabV3Lite with DWASPP |MobileNetV2   |16             |1536x768   |14.48                |73.59     |deeplabv3lite_mobilenetv2_tv                |
-|Cityscapes |FPNPixel2Pixel with DWASPP|MobileNetV2   |32             |1536x768   |**15.37**            |**74.98** |**fpn_pixel2pixel_aspp_mobilenetv2_tv**     |
-
-|Dataset    |Mode Architecture         |Backbone Model|Backbone Stride|Resolution |Complexity (GigaMACS)|MeanIoU%  |Model Configuration Name                    |
-|---------  |----------                |-----------   |-------------- |-----------|--------             |----------|----------------------------------------    |
-|Cityscapes |ERFNet[[4]]               |-             |-              |1024x512   |27.705               |69.7      |-                                           |
-|Cityscapes |SwiftNetMNV2[[5]]         |MobileNetV2   |-              |2048x1024  |41.0                 |75.3      |-                                           |
-|Cityscapes |DeepLabV3Plus[[6,7]]      |MobileNetV2   |16             |           |21.27                |70.71     |-                                           |
-|Cityscapes |DeepLabV3Plus[[6,7]]      |Xception65    |16             |           |418.64               |78.79     |-                                           |
+|Dataset    |Mode Architecture         |Backbone Model |Backbone Stride|Resolution |Complexity (GigaMACS)|MeanIoU%  |Model Configuration Name                  |
+|---------  |----------                |-----------    |-------------- |-----------|--------             |----------|----------------------------------------  |
+|Cityscapes |FPNPixel2Pixel with DWASPP|FD-MobileNetV2 |64             |768x384    |0.99                 |62.43     |fpn_pixel2pixel_aspp_mobilenetv2_tv_fd    |
+|Cityscapes |DeepLabV3Lite with DWASPP |MobileNetV2    |16             |768x384    |**3.54**             |**69.13** |**deeplabv3lite_mobilenetv2_tv**          |
+|Cityscapes |FPNPixel2Pixel            |MobileNetV2    |32(\*2\*2)     |768x384    |3.66                 |70.30     |fpn_pixel2pixel_mobilenetv2_tv            |
+|Cityscapes |FPNPixel2Pixel with DWASPP|MobileNetV2    |32             |768x384    |3.84                 |70.39     |fpn_pixel2pixel_aspp_mobilenetv2_tv       |
+|Cityscapes |FPNPixel2Pixel            |FD-MobileNetV2 |64(\*2\*2)     |1536x768   |3.85                 |69.82     |fpn_pixel2pixel_mobilenetv2_tv_fd         |
+|Cityscapes |FPNPixel2Pixel with DWASPP|FD-MobileNetV2 |64             |1536x768   |**3.96**             |**71.28** |**fpn_pixel2pixel_aspp_mobilenetv2_tv_fd**|
+|Cityscapes |FPNPixel2Pixel with DWASPP|FD-MobileNetV2 |64             |2048x1024  |7.03                 |72.67     |fpn_pixel2pixel_aspp_mobilenetv2_tv_fd    |
+|Cityscapes |DeepLabV3Lite with DWASPP |MobileNetV2    |16             |1536x768   |14.48                |73.59     |deeplabv3lite_mobilenetv2_tv              |
+|Cityscapes |FPNPixel2Pixel with DWASPP|MobileNetV2    |32             |1536x768   |**15.37**            |**74.98** |**fpn_pixel2pixel_aspp_mobilenetv2_tv**   |
+|Cityscapes |FPNPixel2Pixel with DWASPP|FD-ResNet50    |64             |1536x768   |30.91                |-         |fpn_pixel2pixel_aspp_resnet50_fd          |
+|Cityscapes |FPNPixel2Pixel with DWASPP|ResNet50       |32             |1536x768   |114.42               |-         |fpn_pixel2pixel_aspp_resnet50             |
+
+|Dataset    |Mode Architecture         |Backbone Model |Backbone Stride|Resolution |Complexity (GigaMACS)|MeanIoU%  |Model Configuration Name                  |
+|---------  |----------                |-----------    |-------------- |-----------|--------             |----------|----------------------------------------  |
+|Cityscapes |ERFNet[[4]]               |-              |-              |1024x512   |27.705               |69.7      |-                                         |
+|Cityscapes |SwiftNetMNV2[[5]]         |MobileNetV2    |-              |2048x1024  |41.0                 |75.3      |-                                         |
+|Cityscapes |DeepLabV3Plus[[6,7]]      |MobileNetV2    |16             |           |21.27                |70.71     |-                                         |
+|Cityscapes |DeepLabV3Plus[[6,7]]      |Xception65     |16             |           |418.64               |78.79     |-                                         |
 
 Notes: 
 - (\*2\*2) in the above table represents two additional Depthwise Separable Convolutions with strides (at the end of the backbone encoder). 
-- Backbione encoder stride of 64 (some rows in the above table) is achieved by Fast Downsampling Strategy [8]
+- FD-MobileNetV2 Backbone uses a stride of 64 (this is used in some rows of the above table) and is achieved by Fast Downsampling Strategy [8]
 
 ## References
 [1]The Cityscapes Dataset for Semantic Urban Scene Understanding, Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, Bernt Schiele, CVPR 2016, https://www.cityscapes-dataset.com/
index 35884902cd3fe765659159bffc110adf4388b5bc..f447a3d1ed1d8b21fa6e84f1fa2b3432808ce78c 100755 (executable)
 #python ./scripts/train_segmentation_main.py --dataset_name cityscapes_segmentation --model_name fpn_pixel2pixel_aspp_resnet50 --data_path ./data/datasets/cityscapes/data --img_resize 768 1536 --rand_crop 512 1024 --output_size 1024 2048 --gpus 0 1 \
 #--pretrained https://download.pytorch.org/models/resnet50-19c8e357.pth
 
+#### Cityscapes Semantic Segmentation - Training with FD-ResNet50+FPN - Low Complexity Model
+#python ./scripts/train_segmentation_main.py --dataset_name cityscapes_segmentation --model_name fpn_pixel2pixel_aspp_resnet50_fd --data_path ./data/datasets/cityscapes/data --img_resize 768 1536 --rand_crop 512 1024 --output_size 1024 2048 --gpus 0 1 \
+#--pretrained https://download.pytorch.org/models/resnet50-19c8e357.pth
+
+
 #### VOC Segmentation - Training with MobileNetV2+DeeplabV3Lite
 #python ./scripts/train_segmentation_main.py --dataset_name voc_segmentation --model_name deeplabv3lite_mobilenetv2_tv --data_path ./data/datasets/voc --img_resize 512 512 --output_size 512 512 --gpus 0 1 \
 #--pretrained https://download.pytorch.org/models/mobilenet_v2-b0353104.pth