refactor & add lanedet

.gitignore

@@ -11,6 +11,7 @@ __pycache__/
 
 *.egg-info/
 *.egg
+*.eggs
 
 output.txt
 .vscode/*
@@ -19,6 +20,8 @@ tmp.*
 *.pt
 *.pth
 *.un~
+*.so
+build
 
 lane
 debug

README.md

@@ -24,7 +24,7 @@ Supported detectors:
 - [x] [SCNN](configs/scnn)
 - [x] [UFLD](configs/ufld)
 - [x] [RESA](configs/resa)
-- [ ] LaneATT (comming sonn)
+- [x] [LaneATT](configs/laneatt)
 - [ ] LaneAF (comming sonn)
 
 ## Installation
@@ -55,7 +55,7 @@ conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
 pip install torch torchvision
 
 # Install python packages
-pip install -r requirements.txt
+python setup.py build develop
 ```
 
 ### Data preparation
@@ -98,7 +98,7 @@ $TUSIMPLEROOT/test_label.json # test label json file
 For Tusimple, the segmentation annotation is not provided, hence we need to generate segmentation from the json annotation. 
 
 ```Shell
-python scripts/generate_seg_tusimple.py --root $TUSIMPLEROOT
+python tools/generate_seg_tusimple.py --root $TUSIMPLEROOT
 # this will generate seg_label directory
 ```
 
@@ -128,7 +128,13 @@ For example, run
 python main.py configs/resa/resa50_culane.py --validate --load_from culane_resnet50.pth --gpus 0 1 2 3
 ```
 
-For visualization, just add `--view`.
+Currently, this code can output the visualization result when testing, just add `--view`.
+We will get the visualization result in `work_dirs/xxx/xxx/visualization`.
+
+For example, run
+```Shell
+python main.py configs/resa/resa50_culane.py --validate --load_from culane_resnet50.pth --gpus 0 --view
+```
 
 
 ## Contributing
@@ -146,6 +152,7 @@ This project is released under the [Apache 2.0 license](LICNESE).
 * [XingangPan/SCNN](https://github.com/XingangPan/SCNN)
 * [ZJULearning/resa](https://github.com/ZJULearning/resa)
 * [cfzd/Ultra-Fast-Lane-Detection](https://github.com/cfzd/Ultra-Fast-Lane-Detection)
+* [lucastabelini/LaneATT](https://github.com/lucastabelini/LaneATT)
 <!--te-->
 
 <!-- 

configs/laneatt/README.md

@@ -0,0 +1,25 @@
+# Keep your Eyes on the Lane: Real-time Attention-guided Lane Detection
+
+## Introduction
+
+```latex
+@InProceedings{tabelini2021cvpr,
+  author    = {Lucas Tabelini
+               and Rodrigo Berriel
+               and Thiago M. Paix\~ao
+               and Claudine Badue
+               and Alberto Ferreira De Souza
+               and Thiago Oliveira-Santos},
+  title     = {{Keep your Eyes on the Lane: Real-time Attention-guided Lane Detection}},
+  booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
+  year      = {2021}
+}
+```
+
+## Models
+
+| Architecture| Backbone |Dataset | Metric | Config| Checkpoints  |
+|-------------|----------|--------|--------|-------|--------------|
+| LaneATT        | ResNet18 | CULane || [config](https://github.com/Turoad/lanedet/blob/main/configs/laneatt/resnet18_culane.py)  | comming soon|
+| LaneATT        | ResNet18 | TuSimple || [config](https://github.com/Turoad/lanedet/blob/main/configs/laneatt/resnet18_tusimple.py)  | comming soon|
+| LaneATT        | ResNet34 | TuSimple || [config](https://github.com/Turoad/lanedet/blob/main/configs/laneatt/resnet34_tusimple.py)  | comming soon|

configs/laneatt/resnet18_culane.py

@@ -0,0 +1,119 @@
+net = dict(
+    type='Detector',
+)
+
+backbone = dict(
+    type='ResNetWrapper',
+    resnet='resnet18',
+    pretrained=True,
+    replace_stride_with_dilation=[False, False, False],
+    out_conv=False,
+)
+featuremap_out_channel = 512 
+featuremap_out_stride = 32 
+
+aggregator = None
+
+num_points = 72
+max_lanes = 4
+sample_y = range(589, 230, -20)
+
+heads = dict(type='LaneATT',
+        anchors_freq_path='.cache/culane_anchors_freq.pt',
+        topk_anchors=1000)
+
+train_parameters = dict(
+    conf_threshold=None,
+    nms_thres=15.,
+    nms_topk=3000
+)
+test_parameters = dict(
+    conf_threshold=0.5,
+    nms_thres=50,
+    nms_topk=max_lanes
+)
+
+optimizer = dict(
+  type = 'AdamW',
+  lr = 0.0003,
+)
+
+epochs = 15
+batch_size = 8
+total_iter = (88880 // batch_size) * epochs
+scheduler = dict(
+    type = 'CosineAnnealingLR',
+    T_max = total_iter
+)
+
+eval_ep = 1
+save_ep = 1
+
+img_norm = dict(
+    mean=[103.939, 116.779, 123.68],
+    std=[1., 1., 1.]
+)
+ori_img_w=1640
+ori_img_h=590
+img_w=640 
+img_h=360
+cut_height=0
+
+train_process = [
+    dict(type='GenerateLaneLine', 
+        transforms = (
+            dict(
+                name = 'Affine',
+                parameters = dict(
+                    translate_px = dict(
+                        x = (-25, 25),
+                        y = (-10, 10)
+                    ),
+                    rotate=(-6, 6),
+                    scale=(0.85, 1.15)
+                )
+            ),
+            dict(
+                name = 'HorizontalFlip',
+                parameters = dict(
+                    p=0.5
+                ),
+            )
+        ), 
+        wh = (img_w, img_h),
+    ),
+    dict(type='ToTensor', keys=['img', 'lane_line']),
+] 
+
+val_process = [
+    dict(type='GenerateLaneLine'),
+    dict(type='ToTensor', keys=['img']),
+] 
+
+dataset_path = './data/CULane'
+dataset_type = 'CULane'
+dataset = dict(
+    train=dict(
+        type=dataset_type,
+        data_root=dataset_path,
+        split='train',
+        processes=train_process,
+    ),
+    val=dict(
+        type=dataset_type,
+        data_root=dataset_path,
+        split='test',
+        processes=val_process,
+    ),
+    test=dict(
+        type=dataset_type,
+        data_root=dataset_path,
+        split='test',
+        processes=val_process,
+    )
+)
+
+
+workers = 12
+log_interval = 100
+seed=0

configs/laneatt/resnet18_tusimple.py

@@ -0,0 +1,117 @@
+net = dict(
+    type='Detector',
+)
+
+backbone = dict(
+    type='ResNetWrapper',
+    resnet='resnet18',
+    pretrained=True,
+    replace_stride_with_dilation=[False, False, False],
+    out_conv=False,
+)
+featuremap_out_channel = 512 
+featuremap_out_stride = 32 
+
+aggregator = None
+
+num_points = 72
+max_lanes = 5
+sample_y=range(710, 150, -10)
+
+heads = dict(type='LaneATT',
+        anchors_freq_path='.cache/tusimple_anchors_freq.pt',
+        topk_anchors=1000)
+
+train_parameters = dict(
+    conf_threshold=None,
+    nms_thres=15.,
+    nms_topk=3000
+)
+test_parameters = dict(
+    conf_threshold=0.2,
+    nms_thres=45,
+    nms_topk=max_lanes
+)
+
+optimizer = dict(
+  type = 'Adam',
+  lr = 0.0003,
+)
+
+epochs = 100
+batch_size = 8
+total_iter = (3616 // batch_size) * epochs
+scheduler = dict(
+    type = 'CosineAnnealingLR',
+    T_max = total_iter
+)
+
+eval_ep = 1
+save_ep = epochs
+
+ori_img_w=1280
+ori_img_h=720
+img_w=640 
+img_h=360
+cut_height=0
+
+train_process = [
+    dict(type='GenerateLaneLine', 
+        transforms = (
+            dict(
+                name = 'Affine',
+                parameters = dict(
+                    translate_px = dict(
+                        x = (-25, 25),
+                        y = (-10, 10)
+                    ),
+                    rotate=(-6, 6),
+                    scale=(0.85, 1.15)
+                )
+            ),
+            dict(
+                name = 'HorizontalFlip',
+                parameters = dict(
+                    p=0.5
+                ),
+            )
+        ), 
+        wh = (img_w, img_h),
+    ),
+    dict(type='ToTensor', keys=['img', 'lane_line']),
+] 
+
+val_process = [
+    dict(type='GenerateLaneLine', wh=(img_w, img_h)),
+    dict(type='ToTensor', keys=['img']),
+]
+
+dataset_path = './data/tusimple'
+test_json_file = 'data/tusimple/test_label.json'
+dataset_type = 'TuSimple'
+
+dataset = dict(
+    train=dict(
+        type=dataset_type,
+        data_root=dataset_path,
+        split='trainval',
+        processes=train_process,
+    ),
+    val=dict(
+        type=dataset_type,
+        data_root=dataset_path,
+        split='test',
+        processes=val_process,
+    ),
+    test=dict(
+        type=dataset_type,
+        data_root=dataset_path,
+        split='test',
+        processes=val_process,
+    )
+)
+
+
+workers = 12
+log_interval = 100
+seed=0