diff --git a/README.md b/README.md index fdef93b..950f4aa 100644 --- a/README.md +++ b/README.md @@ -6,8 +6,8 @@ This implementation has been tested on the CamVid and Cityscapes datasets. Curre | Dataset | Classes 1 | Input resolution | Batch size | Epochs | Mean IoU (%) | GPU memory (GiB) | Training time (hours)2 | | :------------------------------------------------------------------: | :------------------: | :--------------: | :--------: | :----: | :---------------: | :--------------: | :-------------------------------: | -| [CamVid](http://mi.eng.cam.ac.uk/research/projects/VideoRec/CamVid/) | 11 | 480x360 | 10 | 300 | 51.083 | 4.2 | 1 | -| [Cityscapes](https://www.cityscapes-dataset.com/) | 19 | 1024x512 | 4 | 300 | 59.034 | 5.4 | 20 | +| [CamVid](http://mi.eng.cam.ac.uk/research/projects/VideoRec/CamVid/) | 11 | 480x360 | 10 | 300 | 52.13 | 4.2 | 1 | +| [Cityscapes](https://www.cityscapes-dataset.com/) | 19 | 1024x512 | 4 | 300 | 59.54 | 5.4 | 20 | 1 When referring to the number of classes, the void/unlabeled class is always excluded.
2 These are just for reference. Implementation, datasets, and hardware changes can lead to very different results. Reference hardware: Nvidia GTX 1070 and an AMD Ryzen 5 3600 3.6GHz. You can also train for 100 epochs or so and get similar mean IoU (± 2%).
diff --git a/save/ENet_CamVid/ENet b/save/ENet_CamVid/ENet index b5782c3..2960b91 100644 Binary files a/save/ENet_CamVid/ENet and b/save/ENet_CamVid/ENet differ diff --git a/save/ENet_CamVid/ENet_summary.txt b/save/ENet_CamVid/ENet_summary.txt index f37fa67..c11f412 100644 --- a/save/ENet_CamVid/ENet_summary.txt +++ b/save/ENet_CamVid/ENet_summary.txt @@ -1,7 +1,7 @@ ARGUMENTS batch_size: 10 dataset: camvid -dataset_dir: ../Datasets/CamVid/ +dataset_dir: ../CamVid/ device: cuda epochs: 300 height: 360 @@ -21,5 +21,5 @@ width: 480 workers: 4 BEST VALIDATION -Epoch: 260 -Mean IoU: 0.6407776379897647 +Epoch: 280 +Mean IoU: 0.6518655444842216 diff --git a/save/ENet_Cityscapes/ENet b/save/ENet_Cityscapes/ENet index 44bc91e..44ede15 100644 Binary files a/save/ENet_Cityscapes/ENet and b/save/ENet_Cityscapes/ENet differ diff --git a/save/ENet_Cityscapes/ENet_summary.txt b/save/ENet_Cityscapes/ENet_summary.txt index 41604f5..daadacb 100644 --- a/save/ENet_Cityscapes/ENet_summary.txt +++ b/save/ENet_Cityscapes/ENet_summary.txt @@ -1,7 +1,7 @@ ARGUMENTS batch_size: 4 dataset: cityscapes -dataset_dir: ../Datasets/Cityscapes/1024x512/ +dataset_dir: ../Cityscapes/1024x512/ device: cuda epochs: 300 height: 512 @@ -14,12 +14,12 @@ mode: train name: ENet print_step: False resume: False -save_dir: save/new_cityscapes/ +save_dir: save/cityscapes_2/ weighing: ENet weight_decay: 0.0002 width: 1024 workers: 4 BEST VALIDATION -Epoch: 230 -Mean IoU: 0.5921075030966009 +Epoch: 250 +Mean IoU: 0.5949690267526815 diff --git a/save/README.md b/save/README.md index 8c8d053..72730db 100644 --- a/save/README.md +++ b/save/README.md @@ -2,20 +2,20 @@ | Dataset | Classes 1 | Input resolution | Batch size | Epochs | Mean IoU (%) | GPU memory (GiB) | Training time (hours)2 | | :------------------------------------------------------------------: | :------------------: | :--------------: | :--------: | :----: | :---------------: | :--------------: | :-------------------------------: | -| [CamVid](http://mi.eng.cam.ac.uk/research/projects/VideoRec/CamVid/) | 11 | 480x360 | 10 | 300 | 51.083 | 4.2 | 1 | -| [Cityscapes](https://www.cityscapes-dataset.com/) | 19 | 1024x512 | 4 | 300 | 59.034 | 5.4 | 20 | +| [CamVid](http://mi.eng.cam.ac.uk/research/projects/VideoRec/CamVid/) | 11 | 480x360 | 10 | 300 | 52.13 | 4.2 | 1 | +| [Cityscapes](https://www.cityscapes-dataset.com/) | 19 | 1024x512 | 4 | 300 | 59.54 | 5.4 | 20 | ## Per-class IoU: CamVid3 | | Sky | Building | Pole | Road | Pavement | Tree | Sign Symbol | Fence | Car | Pedestrian | Bicyclist | | :-----: | :---: | :------: | :---: | :---: | :------: | :---: | :---------: | :---: | :---: | :--------: | :-------: | -| IoU (%) | 89.8 | 68.2 | 19.9 | 90.1 | 71.6 | 62.7 | 17.8 | 15.1 | 65.9 | 25.8 | 34.3 | +| IoU (%) | 90.2 | 68.6 | 22.6 | 91.5 | 73.2 | 63.6 | 19.3 | 16.7 | 65.1 | 27.2 | 35.0 | ## Per-class IoU: Cityscapes4 | | Road | Sidewalk | Building | Wall | Fence | Pole | Traffic light | Traffic Sign | Vegetation | Terrain | Sky | Person | Rider | Car | Truck | Bus | Train | Motorcycle | Bicycle | | :-----: | :---: | :------: | :------: | :---: | :---: | :---: | :-----------: | :----------: | :--------: | :-----: | :---: | :----: | :---: | :---: | :---: | :---: | :---: | :--------: | :-----: | -| IoU (%) | 96.1 | 73.9 | 86.0 | 39.5 | 41.8 | 45.6 | 43.6 | 54.9 | 88.1 | 53.5 | 90.1 | 62.5 | 41.4 | 88.3 | 41.3 | 56.9 | 34.6 | 24.1 | 59.6 | +| IoU (%) | 96.1 | 73.3 | 85.8 | 44.1 | 40.5 | 45.3 | 42.5 | 53.9 | 87.9 | 53.5 | 90.1 | 62.3 | 44.3 | 87.6 | 46.6 | 58.2 | 34.8 | 25.8 | 57.9 | 1 When referring to the number of classes, the void/unlabeled class is always excluded.
2 These are just for reference. Implementation, datasets, and hardware changes can lead to very different results. Reference hardware: Nvidia GTX 1070 and an AMD Ryzen 5 3600 3.6GHz. You can also train for 100 epochs or so and get similar mean IoU (± 2%).