Merge pull request #31 from rapzag/main

Add GOOSE Dataset

README.md

@@ -90,7 +90,7 @@ To be able to run our network, please follow those steps for generating processe
 
 ### Dataset Download: 
 
-Please go to [RUGD](http://rugd.vision/) and [RELLIS-3D](https://github.com/unmannedlab/RELLIS-3D/blob/main/README.md#annotated-data) (we use the ID annotation instead of color annotation for RELLIS-3D) officail website to download their data. 
+Please go to [RUGD](http://rugd.vision/) and [RELLIS-3D](https://github.com/unmannedlab/RELLIS-3D/blob/main/README.md#annotated-data) (we use the ID annotation instead of color annotation for RELLIS-3D) officail website to download their data. Download the [GOOSE](https://goose-dataset.de/docs/setup/#download-dataset) dataset, and extract all zips in `data/goose` to match the below structure.
 <!-- Please structure the downloaded data as follows: -->
 
 Note: Since RELLIS-3D dataset has been updated, please run `python ./tools/process_rellis.py` to remove "pylon_camera_node/", "pylon_camera_node_label_id/" folder, after structure the data as follows:
@@ -117,6 +117,18 @@ GANav
 │   │   │   ├── creek & park-1/2/8 & trail-(1 & 3-7 & 9-15) & village
 │   │   │── RUGD_frames-with-annotations
 │   │   │   ├── creek & park-1/2/8 & trail-(1 & 3-7 & 9-15) & village
+│   ├── goose
+│   │   ├── goose_label_mapping.csv
+│   │   ├── images
+│   │   │   ├── train
+│   │   │   └── val
+│   │   ├── labels
+│   │   │   ├── train
+│   │   │   └── val
+│   │   ├── LICENSE
+│   │   ├── test.txt
+│   │   ├── train.txt
+│   │   └── val.txt
 ├── configs
 ├── tools
 ...
@@ -140,6 +152,12 @@ For RUGD dataset, run:
 
 Replease [x] with 4 or 6, to generated data with 4 annotation groups or 6 annotation groups.
 
+For the GOOSE dataset, run:
+
+   ```
+   python ./tools/convert_datasets/goose_relabel6.py
+   ``` 
+
 ## Training
 
 To train a model on RUGD datasets with our methods on 6 groups:
@@ -154,6 +172,12 @@ To train a model on multiple GPUs(RUGD, 6 groups, 2 GPUs):
 ./tools/dist_train.sh ./configs/ours/ganav_group6_rugd.py 2
 ```
 
+To train on the GOOSE dataset run:
+
+```
+python ./tools/train.py ./configs/ours/ganav_group6_goose.py
+```
+
 ## Testing
 
 An example to evaluate our method with 6 groups on RUGD datasets with mIoU metrics:

configs/_base_/datasets/goose_group6.py

@@ -0,0 +1,62 @@
+# dataset settings
+dataset_type = 'GOOSEDataset_Group6'
+data_root = 'data/goose/'
+img_norm_cfg = dict(
+    mean=[80.66, 77.73, 76.87], std=[61.40, 61.08, 64.47], to_rgb=True)
+img_size = (2048, 1000)
+crop_size = (375, 600)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(type='Resize', img_scale=img_size, ratio_range=(0.5, 2.0)),
+    dict(type='RandomCrop', crop_size=crop_size),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='Normalize', **img_norm_cfg),
+    dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+    dict(type='DefaultFormatBundle'),
+    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(
+        type='MultiScaleFlipAug',
+        img_scale=crop_size,
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
+        flip=False,
+        transforms=[
+            dict(type='Resize', keep_ratio=True),
+            dict(type='RandomFlip'),
+            dict(type='Normalize', **img_norm_cfg),
+            dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
+            dict(type='ImageToTensor', keys=['img']),
+            dict(type='Collect', keys=['img']),
+        ])
+]
+data = dict(
+    samples_per_gpu=1,
+    workers_per_gpu=2,
+    img_size=crop_size,
+    train=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/train/',
+        ann_dir='labels/train/',
+        split='train.txt',
+        pipeline=train_pipeline),
+    val=dict(
+        type=dataset_type,
+        data_root=data_root,
+        img_dir='images/val/',
+        ann_dir='labels/val/',
+        split='val.txt',
+        pipeline=test_pipeline),
+    # test=dict(
+    #     type=dataset_type,
+    #     data_root=data_root,
+    #     img_dir='image',
+    #     ann_dir='labels',
+    #     split='test.txt',
+    #     pipeline=test_pipeline)
+    )

configs/ours/ganav_group6_goose.py

@@ -0,0 +1,27 @@
+_base_ = [
+    '../_base_/models/ours_class_att.py', '../_base_/datasets/goose_group6.py',
+    '../_base_/default_runtime.py'
+]
+
+
+optimizer = dict(type='SGD', lr=0.003, momentum=0.9, weight_decay=4e-5)
+optimizer_config = dict()
+# learning policy
+# runtime settings
+runner = dict(type='IterBasedRunner', max_iters=240000)
+total_iters = 240000
+checkpoint_config = dict(by_epoch=False, interval=16000)
+evaluation = dict(interval=240000, metric='mIoU')
+
+# optimizer
+lr_config = dict(policy='poly', power=0.9, min_lr=1e-4, warmup='linear',
+    warmup_iters=1500,
+    warmup_ratio=1e-6,
+    by_epoch=False)
+
+model=dict(
+    decode_head=dict(img_size=(375, 600)))
+
+data = dict(
+    samples_per_gpu=2,
+    workers_per_gpu=2)

data/goose/.gitignore

@@ -0,0 +1,4 @@
+goose_label_mapping.csv
+images/*
+labels/*
+LICENSE