BridgeQA: Bridging the Gap between 2D and 3D Visual Question Answering: A Fusion Approach for 3D VQA - Official Codebase

This work is accepeted by AAAI 2024.

Installation

Please follow the procedure in INSTALLATION.

Data Preparation

Please follow the same procedure from DATASET. Also refer to ScanQA and ScanRefer.

Results

ScanQA

SQA

Training

All model outputs and checkpoints will be saved under ./outputs/ path. You can find the checkpoints and logs of each run after training. We also provide pretrained or pre-converted files HERE

Quesiton-Conditional View Selection

Question-Declaration Transform

To transform question to corresponding declaration, run following command:

export OPENAI_API_KEY = <your-openai-key>
python compose_decl_from_qa.py --output_qa_file <path/to/decl_file>

The output JSON file of declarations is saved as specfied in output_qa_file option. Replication note: since OpenAI will deprecate its older version GPT-4 of gpt-3.5-0301, and the randomness of nucleus sampling, you might not be able to acquire the same result declaration as ours. You can refer to our pre-converted file that is used in our reported performance.

View Selection

To select views for questions, run following command:

python eval_scene_best_views.py \
    --outfile <path/to/result>  --topk_images 1 \
    --dset_views_path <path/to/views_folder> --nocheck_blank --split "train,val,test_w_obj,test_wo_obj"  \
    --use_composed_qa --composed_qa_json <path/to/composed_decl> \

The result .pkl file (written to specified outfile option) is used for later training procedure as the --i2tfile parameter. You can also use the original question to find the best view at inference time. For SQA, replace the split option to "train,val,test".

Pretraining Detector

To pretrain a VoteNet detector, simply run following command without 2D VLM and QA-related losses:

export PORT=$(shuf -i 2000-3000 -n 1)
export SLURM_GPUS=4
torchrun --nproc_per_node=$SLURM_GPUS --nnodes=1 --rdzv_backend=c10d --rdzv_endpoint=localhost:$PORT \
    scripts/train.py --ddp --use_color --tag detection --optim_name adamw --image_size 512 \
    --use_multiview \
    --dset_views_path <path/to/views_folder> \
    --i2tfile <path/to/i2tfile> \
    --train_batch_size 16 --val_batch_size 64 --lr_blip "5e-5" --wd_blip "0.0" --lr "5e-4" --lr_decay_rate 0.2 \
    --val_step 200 --scene_feature_type full \
    --lr_decay_step 15 35 --val_step 200 --scheduler_type step --lr_blip 5e-5 --wd_blip 0.0 --lr 5e-4 \
    --stage "DET" --cur_criterion "loss" --no_reference

dset_views_path should be set to the dataset frames path where you download in Data Preparation. i2tfile should be set to the .pkl file you obtained at the View Selection step. The training checkpoints, logs and configs will be saved in a new directory with training date-time info under ./outputs. We simply take the last checkpoint model_last.pth as the choice for later VQA training.

NOTE: wandb is used to track the training statistics, if you want to disable it, simply set WANDB_MODE=disabled

Training

To train the VQA model, simply run following command:

export PORT=$(shuf -i 2000-3000 -n 1)
export SLURM_GPUS=8
torchrun --nproc_per_node=$SLURM_GPUS --nnodes=1 --rdzv_backend=c10d --rdzv_endpoint=localhost:$PORT \
    scripts/train.py --ddp --use_color --tag allanswer --optim_name adamw --val_step 500 --image_size 512 --scheduler_type step \
    --use_multiview --use_blip \
    --dset_views_path <path/to/views_folder> \
    --train_batch_size 16 --val_batch_size 2 --lr_blip "1e-5" --wd_blip "0.0" --lr "5e-4"  \
    --val_step 200 --scene_feature_type full \
    --stage VQA --answer_loss_weight 3.0 \
    --i2tfile <path/to/i2tfile> \
    --first_stage_ckpt_path <path/to/detector_ckpt> \
    --use_text_decoder --share_decoder \
    --scene_feature_position paralleltwin --lr_blip3d "3e-5" --scheduler_type step_except_2d \
    --epoch 10 --lr_decay_step 5 8 --lr_decay_step_2d 3 5 7

dset_views_path should be set to the dataset frames path where you download in Data Preparation. i2tfile should be set to the .pkl file you obtained at the View Selection step. first_stage_ckpt_path should be set to the model result folder under ./outputs at detector pretrain stage. The training checkpoints, logs and configs will be saved in a new directory with training date-time info under ./outputs.

Inference

To inference and make a prediction, simply run following command:

export PORT=$(shuf -i 2000-3000 -n 1)
torchrun --nproc_per_node=$SLURM_GPUS --nnodes=1 --rdzv_backend=c10d --rdzv_endpoint=localhost:$PORT \
    scripts/predict.py \
    --folder <path/to/traininin_output> \
    --i2tfile <path/to/i2tfile> \
    --test_type <split-to-test> --batch_size 2 \

and the prediction can be found at the same folder as the training output. We use the best checkpoint (the model.pth) verified on validation set to predict the results on test splits.

Checkpoints and Pre-converted files

The converted declaration file is generated by gpt-3.5-0301.

TODO

• [x] Make copy of BLIP codes
• [x] Clean-up model codes
• [x] Clean-up training codes
• [x] Test training
• [x] Clean-up prediction codes
• [x] Test prediction
• [x] Clean-up q2d codes
• [x] Test q2d codes
• [x] Clean-up image-question selection codes
• [x] Test image-question selection codes
• [x] Clean-up detector pre-training.
• [x] Test detector pre-training.
• [x] Clean-up dependencies.
• [x] Report performance with this cleaned implementation
• [x] Update view-selection, training instructions
• [x] Update evaluation instructions
• [x] Update q2d instructions
• [x] Upload pretrained checkpoints and i2t mappings for ScanQA
• [ ] Add and combine SQA3D training codes
• [ ] Upload pretrained checkpoints and i2t mappings for SQA

Acknowledgements

We would like to thank facebookresearch/votenet for the 3D object detection, daveredrum/ScanRefer for the 3D localization codebase and ScanQA for 3D question answering codebase. We also thank BLIP for the 2D Vision-Language Model and architecture codebase.

License

BridgeQA is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License.

Copyright (c) Wentao Mo, 2024.