Forward Direction in visualize_sub_traj.py

Aerial Vision-and-Dialog Navigation

The ability to converse with humans and follow natural language commands is crucial for intelligent unmanned aerial vehicles (a.k.a. drones). It can relieve people's burden of holding a controller all the time, allow multitasking, and make drone control more accessible for people with disabilities or with their hands occupied. To this end, we introduce Aerial Vision-and-Dialog Navigation (AVDN), to navigate a drone via natural language conversation. We build a drone simulator with a continuous photorealistic environment and collect a new AVDN dataset of over 3k recorded navigation trajectories with asynchronous human-human dialogs between commanders and followers. The commander provides initial navigation instruction and further guidance by request, while the follower navigates the drone in the simulator and asks questions when needed. During data collection, followers' attention on the drone's visual observation is also recorded. Based on the AVDN dataset, we study the tasks of aerial navigation from (full) dialog history and propose an effective Human Attention Aided Transformer model (HAA-Transformer), which learns to predict both navigation waypoints and human attention.

Todos:

AVDN Challenge and Leaderboard

Based on the AVDN dataset, we are hosting an ICCV 2023 Challenge (co-located at the ICCV 2023 CLVL workshop) for the Aerial Navigation from Dialog History (ANDH) task on Eval.ai: https://eval.ai/web/challenges/challenge-page/2049/overview

Download Data

Download xView data

Our AVDN dataset uses satellite images from the xView dataset. Follow the instruction at https://challenge.xviewdataset.org/data-download to download xView dataset.

Then move the images in xView dataset to under AVDN directory. (Assume the xView images are at ./XVIEW_images):

mkdir -p Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/train_images

cp -r XVIEW_images/*.tif Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/train_images/

Download AVDN datasets

(https://sites.google.com/view/aerial-vision-and-dialog/home):

mkdir -p Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/annotations

gdown 1xUHnrYaNGe_IBG7W1ecaf6U2cyuBfYLr -O Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/annotations/train_data.json

gdown 1mtT3AVJQNEbjKkH6aINX3kj7ROADkBET -O Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/annotations/val_seen_data.json

gdown 17fVSHmuB3EFHkfNRZle6kgVcvZcumsJr -O Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/annotations/val_unseen_data.json

gdown 14BijI07ukKCSDh3T_RmUG83z6Oa75M-U -O Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/annotations/test_unseen_data.json

Training and Evaluation

Download pre-trained xview-yolov3 weights and configuration file

mkdir -p Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/pretrain_weights

gdown 1Ke-pA5jpq1-fsEwAch_iRCtJHx6rQc-Z -O Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/pretrain_weights/best.pt

gdown 1n6RMWcHAbS6DA7BBug6n5dyN6NPjiPjh -O Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/pretrain_weights/yolo_v3.cfg

Download the training checkpoints corresponding to the experiments in the AVDN paper

mkdir -p Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/et_haa/ckpts/

mkdir -p Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/lstm_haa/ckpts/

gdown 1fA6ckLVA-gsiOmWmOMkqJggTLbiJpFBI -O Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/et_haa/ckpts/best_val_unseen

gdown 1RYjo_vc5m5ZRUcjIFojZjke8RhlfX90I -O Aerial-Vision-and-Dialog-Navigation/datasets/AVDN/lstm_haa/ckpts/best_val_unseen

Install requirements

pip install torch==1.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html

pip install torchvision==0.12.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html

pip install -r requirements.txt

Run training or evaluation:

The script, scripts/avdn_paper/run_et_haa.sh, includes commands for train and evaluate Human Attention Aided Transformer (HAA-Transformer) model.

The script, scripts/avdn_paper/run_lstm_haa.sh, includes commands for train and evaluate Human Attention Aided LSTM (HAA-LSTM) model.

cd Aerial-Vision-and-Dialog-Navigation/src

# For Human Attention Aided Transformer model
bash scripts/avdn_paper/run_et_haa.sh 

# For Human Attention Aided LSTM model
bash scripts/avdn_paper/run_lstm_haa.sh

If you find this useful, please cite

@inproceedings{fan-etal-2023-aerial,
    title = "Aerial Vision-and-Dialog Navigation",
    author = "Fan, Yue  and
      Chen, Winson  and
      Jiang, Tongzhou  and
      Zhou, Chun  and
      Zhang, Yi  and
      Wang, Xin Eric",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2023",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.findings-acl.190",
    doi = "10.18653/v1/2023.findings-acl.190",
    pages = "3043--3061",
}

	_net_next_pos = 1e5*(teacher_a[i][0] - current_pos)
	_net_y = np.round(1e5*((corners[i][0] + corners[i][1])/2 - current_pos)).astype(np.int)
	_net_x = np.round(1e5*((corners[i][1] + corners[i][2])/2 - current_pos)).astype(np.int)

	A = np.mat([[_net_x[0],_net_y[0]],[_net_x[1],_net_y[1]]])
	b = np.mat([_net_next_pos[0],_net_next_pos[1]]).T
	r = np.linalg.solve(A,b)

	gt_next_pos_ratio = [r[0,0], r[1,0]]

	scores['gp'] = gt_net_lengths - \
	np.linalg.norm(path[-1] - gt_path[-2])11.131e4
	scores['oracle_gp'] = gt_net_lengths - \
	np.min([np.linalg.norm(path[x] - gt_path[-1]) for x in range(len(path)) ])11.131e4

	def gps_to_img_coords(self, gps, ob):

	gps_botm_left = ob['gps_botm_left']
	gps_top_right = ob['gps_top_right']
	lng_ratio = ob['lng_ratio']
	lat_ratio = ob['lat_ratio']

	return int(round((gps[1] - gps_botm_left[1]) / lat_ratio)), int(round((gps_top_right[0] - gps[0]) / lat_ratio))

	# Compute loss
	for i in range(len(obs)):
	# if the function teacher_action determins that the current view is the final position, no action should be made
	if type(target[i][0]) != type(-100):
	cuda_gt_next_pos_ratio = torch.from_numpy(target[i][0]).cuda()
	ml_loss += self.progress_regression(pred_next_pos_ratio[i,:], cuda_gt_next_pos_ratio)
	ml_loss += self.progress_regression((torch.atan2(pred_next_pos_ratio[i,0], pred_next_pos_ratio[i,1]+1e-5*np.random.rand(1)[0]) /3.14159 + 2) / 2 %1 ,

	teacher_a[i][1] = float((np.linalg.norm(obs[i]['gt_path_corners'][closest_step_index][0] -\
	obs[i]['gt_path_corners'][closest_step_index][1])\
	11.131e4 -40 )/ (400-40))

	if ended[i] or progress[i] > 0.5:
	teacher_a[i][0] = np.array([0,0], dtype=np.float32)
	continue

eric-ai-lab / aerial-vision-and-dialog-navigation Goto Github PK

aerial-vision-and-dialog-navigation's Introduction

Aerial Vision-and-Dialog Navigation

AVDN Challenge and Leaderboard

Download Data

Training and Evaluation

If you find this useful, please cite

aerial-vision-and-dialog-navigation's People

Contributors

Stargazers

Watchers

Forkers

aerial-vision-and-dialog-navigation's Issues

Recommend Projects

Recommend Topics

Recommend Org