Dronescapes dataset

As introduced in our ICCV 2023 workshop paper: link

Note: We are extending this dataset in another repository. GT data for benchmarking is the same, but we are generating modalities as inputs: dronescapes-2024.

1. Downloading the data

Option 1. Download the pre-processed dataset from HuggingFace repository

git lfs install # Make sure you have git-lfs installed (https://git-lfs.com)
git clone https://huggingface.co/datasets/Meehai/dronescapes

Note: the dataset has about 200GB, so it may take a while to clone it.

2. Using the data

As per the split from the paper:

Split

The data is in data/* (see the ls call above, it should match even if you download from huggingface).

2.1 Using the provided viewer

The simplest way to explore the data is to use the provided notebook. Upon running it, you should get a collage with all the default tasks, like the picture at the top.

For a CLI-only method, you can use the provided reader as well:

python scripts/dronescapes_viewer.py data/test_set_annotated_only/ # or any of the 8 directories in data/

Expected output

[MultiTaskDataset]
 - Path: '/export/home/proiecte/aux/mihai_cristian.pirvu/datasets/dronescapes/data/test_set_annotated_only'
 - Tasks (11): [DepthRepresentation(depth_dpt), DepthRepresentation(depth_sfm_manual202204), DepthRepresentation(depth_ufo), ColorRepresentation(edges_dexined), EdgesRepresentation(edges_gb), NpzRepresentation(normals_sfm_manual202204), OpticalFlowRepresentation(opticalflow_rife), ColorRepresentation(rgb), SemanticRepresentation(semantic_mask2former_swin_mapillary_converted), SemanticRepresentation(semantic_segprop8), ColorRepresentation(softseg_gb)]
 - Length: 116
 - Handle missing data mode: 'fill_none'
== Shapes ==
{'depth_dpt': torch.Size([540, 960]),
 'depth_sfm_manual202204': torch.Size([540, 960]),
 'depth_ufo': torch.Size([540, 960, 1]),
 'edges_dexined': torch.Size([540, 960]),
 'edges_gb': torch.Size([540, 960, 1]),
 'normals_sfm_manual202204': torch.Size([540, 960, 3]),
 'opticalflow_rife': torch.Size([540, 960, 2]),
 'rgb': torch.Size([540, 960, 3]),
 'semantic_mask2former_swin_mapillary_converted': torch.Size([540, 960, 8]),
 'semantic_segprop8': torch.Size([540, 960, 8]),
 'softseg_gb': torch.Size([540, 960, 3])}
== Random loaded item ==
{'depth_dpt': tensor[540, 960] n=518400 (2.0Mb) x∈[0.043, 1.000] μ=0.341 σ=0.418,
 'depth_sfm_manual202204': None,
 'depth_ufo': tensor[540, 960, 1] n=518400 (2.0Mb) x∈[0.115, 0.588] μ=0.297 σ=0.138,
 'edges_dexined': tensor[540, 960] n=518400 (2.0Mb) x∈[0.000, 0.004] μ=0.003 σ=0.001,
 'edges_gb': tensor[540, 960, 1] n=518400 (2.0Mb) x∈[0., 1.000] μ=0.063 σ=0.100,
 'normals_sfm_manual202204': None,
 'opticalflow_rife': tensor[540, 960, 2] n=1036800 (4.0Mb) x∈[-0.004, 0.005] μ=0.000 σ=0.000,
 'rgb': tensor[540, 960, 3] n=1555200 (5.9Mb) x∈[0., 1.000] μ=0.392 σ=0.238,
 'semantic_mask2former_swin_mapillary_converted': tensor[540, 960, 8] n=4147200 (16Mb) x∈[0., 1.000] μ=0.125 σ=0.331,
 'semantic_segprop8': tensor[540, 960, 8] n=4147200 (16Mb) x∈[0., 1.000] μ=0.125 σ=0.331,
 'softseg_gb': tensor[540, 960, 3] n=1555200 (5.9Mb) x∈[0., 0.004] μ=0.002 σ=0.001}
== Random loaded batch ==
{'depth_dpt': tensor[5, 540, 960] n=2592000 (9.9Mb) x∈[0.043, 1.000] μ=0.340 σ=0.417,
 'depth_sfm_manual202204': tensor[5, 540, 960] n=2592000 (9.9Mb) NaN!,
 'depth_ufo': tensor[5, 540, 960, 1] n=2592000 (9.9Mb) x∈[0.115, 0.588] μ=0.296 σ=0.137,
 'edges_dexined': tensor[5, 540, 960] n=2592000 (9.9Mb) x∈[0.000, 0.004] μ=0.003 σ=0.001,
 'edges_gb': tensor[5, 540, 960, 1] n=2592000 (9.9Mb) x∈[0., 1.000] μ=0.063 σ=0.102,
 'normals_sfm_manual202204': tensor[5, 540, 960, 3] n=7776000 (30Mb) NaN!,
 'opticalflow_rife': tensor[5, 540, 960, 2] n=5184000 (20Mb) x∈[-0.004, 0.006] μ=0.000 σ=0.000,
 'rgb': tensor[5, 540, 960, 3] n=7776000 (30Mb) x∈[0., 1.000] μ=0.393 σ=0.238,
 'semantic_mask2former_swin_mapillary_converted': tensor[5, 540, 960, 8] n=20736000 (79Mb) x∈[0., 1.000] μ=0.125 σ=0.331,
 'semantic_segprop8': tensor[5, 540, 960, 8] n=20736000 (79Mb) x∈[0., 1.000] μ=0.125 σ=0.331,
 'softseg_gb': tensor[5, 540, 960, 3] n=7776000 (30Mb) x∈[0., 0.004] μ=0.002 σ=0.001}
== Random loaded batch using torch DataLoader ==
{'depth_dpt': tensor[5, 540, 960] n=2592000 (9.9Mb) x∈[0.025, 1.000] μ=0.216 σ=0.343,
 'depth_sfm_manual202204': tensor[5, 540, 960] n=2592000 (9.9Mb) x∈[0., 1.000] μ=0.562 σ=0.335 NaN!,
 'depth_ufo': tensor[5, 540, 960, 1] n=2592000 (9.9Mb) x∈[0.100, 0.580] μ=0.290 σ=0.128,
 'edges_dexined': tensor[5, 540, 960] n=2592000 (9.9Mb) x∈[0.000, 0.004] μ=0.003 σ=0.001,
 'edges_gb': tensor[5, 540, 960, 1] n=2592000 (9.9Mb) x∈[0., 1.000] μ=0.079 σ=0.116,
 'normals_sfm_manual202204': tensor[5, 540, 960, 3] n=7776000 (30Mb) x∈[0.000, 1.000] μ=0.552 σ=0.253 NaN!,
 'opticalflow_rife': tensor[5, 540, 960, 2] n=5184000 (20Mb) x∈[-0.013, 0.016] μ=0.000 σ=0.004,
 'rgb': tensor[5, 540, 960, 3] n=7776000 (30Mb) x∈[0., 1.000] μ=0.338 σ=0.237,
 'semantic_mask2former_swin_mapillary_converted': tensor[5, 540, 960, 8] n=20736000 (79Mb) x∈[0., 1.000] μ=0.125 σ=0.331,
 'semantic_segprop8': tensor[5, 540, 960, 8] n=20736000 (79Mb) x∈[0., 1.000] μ=0.125 σ=0.331,
 'softseg_gb': tensor[5, 540, 960, 3] n=7776000 (30Mb) x∈[0., 0.004] μ=0.002 σ=0.001}

3. Evaluation for semantic segmentation

We evaluate in the paper on the 3 test scenes (unsees at train) as well as the semi-supervised scenes (seen, but different split) against the human annotated frames. The general evaluation script is in scripts/evaluate_semantic_segmentation.py.

General usage is:

python scripts/evaluate_semantic_segmentation.py y_dir gt_dir -o results.csv --classes C1 C2 .. Cn
[--class_weights W1 W2 ... Wn] [--scenes s1 s2 ... sm]

Script explanation

The script is a bit convoluted, so let's break it into parts:

y_dir and gt_dir Two directories of .npz files in the same format as the dataset (y_dir/1.npz, gt_dir/55.npz etc.)
classes A list of classes in the order that they appear in the predictions and gt files
class_weights (optional, but used in paper) How much to weigh each class. In the paper we compute these weights as the number of pixels in all the dataset (train/val/semisup/test) for each of the 8 classes resulting in the numbers below.
scenes if the y_dir and gt_dir contains multiple scenes that you want to evaluate separately, the script allows you to pass the prefix of all the scenes. For example, in data/test_set_annotated_only/semantic_segprop8/ there are actually 3 scenes in the npz files and in the paper, we evaluate each scene independently. Even though the script outputs one csv file with predictions for each npz file, the scenes are used for proper aggregation at scene level.

Reproducing paper results for Mask2Former

python scripts/evaluate_semantic_segmentation.py \
  data/test_set_annotated_only/semantic_mask2former_swin_mapillary_converted/ \
  data/test_set_annotated_only/semantic_segprop8/ \
  -o results.csv \
  --classes land forest residential road little-objects water sky hill \
  --class_weights 0.28172092 0.30589653 0.13341699 0.05937348 0.00474491 0.05987466 0.08660721 0.06836531 \
  --scenes barsana_DJI_0500_0501_combined_sliced_2700_14700 comana_DJI_0881_full norway_210821_DJI_0015_full

Should output:

scene                                             iou      f1                                                           
barsana_DJI_0500_0501_combined_sliced_2700_14700  63.371  75.338
comana_DJI_0881_full                              60.559  73.779
norway_210821_DJI_0015_full                       37.986  45.939
mean                                              53.972  65.019

Not providing --scenes will make an average across all 3 scenes (not average after each metric individually):

          iou      f1
scene
all    60.456  73.261

3.1 Official benchmark

IoU

method	#paramters	average	barsana_DJI_0500_0501_combined_sliced_2700_14700	comana_DJI_0881_full	norway_210821_DJI_0015_full
Mask2Former	216M	53.97	63.37	60.55	37.98
NGC(LR)	32M	40.75	46.51	45.59	30.17
CShift[^1]	n/a	39.67	46.27	43.67	29.09
NGC[^1]	32M	35.32	44.34	38.99	22.63
SafeUAV[^1]	1.1M	32.79	n/a	n/a	n/a

[^1]: reported in the Dronescapes paper.

F1 Score

method	#paramters	average	barsana_DJI_0500_0501_combined_sliced_2700_14700	comana_DJI_0881_full	norway_210821_DJI_0015_full
Mask2Former	216M	65.01	75.33	73.77	45.93