MT-EQA

Multi-Target Embodied Question Answering

Introduction

We present a generalization of EQA -- Multi-Target EQA (MT-EQA). Specifically, we study questions that have multiple targets in them, such as Is the dresser in the bedroom bigger than the oven in the kitchen?", where the agent has to navigate to multiple locations (dresser in bedroom", oven in kitchen") and perform comparative reasoning (dresser" bigger than ``oven") before it can answer a question. Such questions require the development of entirely new modules or components in the agent. To address this, we propose a modular architecture composed of a program generator, a controller, a navigator, and a VQA module. The program generator converts the given question into sequential executable sub-programs; the navigator guides the agent to multiple locations pertinent to the navigation-related sub-programs; and the controller learns to select relevant observations along its path. These observations are then fed to the VQA module to predict the answer.

Citation

@inproceedings{yu2019mteqa,
  title={Multi-Target Embodied Question Answering},
  author={Yu, Licheng and Chen, Xinlei and Gkioxari, Georgia and Bansal, Mohit and Berg, Tamara L and Batra, Dhruv},
  booktitle={CVPR},
  year={2019}
}

Data Generation

Go to eqa_data folder and do the followings:

1. How to generate question-answer pairs

• Run ./scripts/run_question_generation.sh, this would save qas to cache/question-gen-outputs/questions_pruned_v2.json.
• Run python question-gen/make_object_id_to_cat.py to make mapping from hid.oid to fine_class.

2. Generate graphs, connMaps, and shortest-paths

• Run python shortest-path-gen/generate-conn-maps.py to build graph and connMaps. This would:
  • save graphs to cache/3d-graphs.
  • save connMaps to cache/target-obj-conn-maps.
  • update qas to cache/question-gen-outputs/questions_pruned_mt_with_conn.json.
• Run python shortest-path-gen/get_best_view.py saving best-view points and ious. Note the points information [x, 1.0, z, yaw] is in coordinate system (not grid system).
• Run python shortest-path-gen/add_program_mt.py to add parsed programs to questions, update qas to cache/question-gen-outputs/questions_pruned_mt_with_conn_program.json.
• Run python shortest-path-gen/generate-bestview-paths-mt.py to compute shortest paths connecting start, best-view, end points. Note we intentionally make a faked end point for testing if agent learns to stop at right position (before the faked end point).
• Run python shortest-path-gen/filter_questions_mt.py to filter out those questions without shortest paths or of low entropy, saving filtered questions to questions_mt_paths_nearby_source_best_view_program.json.

3. For installing House3D

• Follow the instructions[https://github.com/facebookresearch/House3D].

Imitation Learning for Nav+Ctrl+cVQA

Go to nav_loc_vqa folder and do the followings:

1. Prepare House Data (conn-maps, graphs, shortest-paths, images, features, etc)

• Copy or Symlink ../eqa_data/cache/question-gen-outputs, ../eqa_data/cache/shortest-paths-mt, ../eqa_data/target-obj-bestview-pos and ../eqa_data/target-obj-conn-maps to ./data
• Run ./tools/generate_path_imgs.py and ./tools/generate_path_cube_imgs.py to extract 1st-person images along paths.
• Run ./tools/extract_feats.py and ./tools/extract_cube_feats.py to extract 1st-person features along paths.
• Run ./tools/compute_meta_info.py to compute room meta info.

2. Train and Eval IL

• Run ./tools/prepro_imitation_data.py and ./tools/prepro_reinforce_data.py
• Run ./tools/eval_gd_path.py to get results on ground-truth paths.

3. Evaluate RL-finetuned Model (after checking eqa_nav)

• Run ./tools/evaleqa_nr.py

Reinforcement Learning Finetuning for Navigators

Go to eqa_nav folder and do the followings:

1. Prepare House Data (conn-maps, graphs, shortest-paths, images, features, etc)

• Copy or Symlink ../eqa_data/cache/question-gen-outputs, ../eqa_data/cache/shortest-paths-mt, ../eqa_data/target-obj-bestview-pos and ../eqa_data/target-obj-conn-maps to ./data
• Copy or Symlink ../nav_loc_eqa/cache/path_feats, ../nav_loc_eqa/cache/path_images and ../nav_loc_eqa/cache/path_to_room_meta to ./cache.

2. Prepare Navigation Data

• Run ./tools/prepro_im.py for imitation learning data.
• Run ./tools/prepro_rl.py for reinforcement finetuning data.

3. Train IL-based room-navigator and object-navigator

• Run ./tools/train_im_object.py and ./tools/train_im_room.py for imitation learning.
• Run ./tools/eval_im_object.py and ./tools/eval_im_room.py for imitation learning evaluation.

4. Finetune using RL

• Run ./tools/train_rl_object.py and ./tools/train_rl_room.py for reinforcement finetuning.
• Run ./tools/eval_nav_object.py and ./tools/eval_nav_room.py for navigation evaluation.

Contact

This project is maintained by Licheng Yu.

License

BSD