bqm1111 / odin

Code for the paper: "ODIN: A Single Model for 2D and 3D Perception" (CVPR 2024)

ODIN: A Single Model for 2D and 3D Segmentation

Authors: Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios, Adam W. Harley, Gabriel Sarch, Kriti Aggarwal, Vishrav Chaudhary, Katerina Fragkiadaki.

Official implementation of "ODIN: A Single Model for 2D and 3D Segmentation", CVPR 2024 (Highlight).

Installation

Make sure you are using a newer version of GCC>=9.2.0

export TORCH_CUDA_ARCH_LIST="6.0 6.1 6.2 7.0 7.2 7.5 8.0 8.6"
conda create -n odin python=3.10
conda activate odin
pip install torch==2.2.0+cu118 torchvision==0.17.0+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
pip install torch-scatter -f https://data.pyg.org/whl/torch-2.2.0+cu118.html
pip install -r requirements.txt
sh init.sh

Data Preparation

Please refer to README in data_preparation folder for individual datasets. For eg. ScanNet data preparation README

Usage

We provide training scripts for various datasets in scripts folder. Please refer to these scripts for training ODIN.

• Modify DETECTRON2_DATASETS to the path where you store the Posed RGB-D data.
• Modify MODEL.WEIGHTS to load pre-trained weights. We use weights from Mask2Former. You can download the Mask2Former-ResNet and Mask2Former-Swin Weights. MODEL.WEIGHTS can also accept link to the checkpoint as well, so you can directly supply these links as the argument value.
• ODIN Pre-trained weights are provided below in the Model-Zoo. Simply point to these weights using the MODEL.WEIGHTS to run inference. You would also need to add --eval-only flag for running evaluation.
• SOLVER.IMS_PER_BATCH controls the batch size. This is effective batch size i.e. if you are running on 2 GPUs and the batch size is set to 6, you are using bs=3 per GPU.
• SOLVER.TEST_IMS_PER_BATCH controls the (effective) test batch size. Since, there are variable number of images in a scene, we use bs=1 per GPU at test time. MAX_FRAME_NUM=-1 means that it loads all images in a scene for inference, which is our usual strategy. In some datasets, the images can simply be too large, thus there we actually set a maximum limit on images.
• INPUT.SAMPLING_FRAME_NUM controls the number of images we sample at test time -- for eg. in ScanNet, we train on 25 image chunks at training time.
• CHECKPOINT_PERIOD is the number of iterations after which a checkpoint is saved. EVAL_PERIOD specifies the number of steps after which the eval is run.
• OUTPUT_DIR stores the checkpoints and the tensorboard logs. --resume resumes the training from the last checkpoint stored in OUTPUT_DIR. If no checkpoint is present, it loads the weights from MODEL.WEIGHTS

Model Zoo

ScanNet Instance Segmentation

Dataset	mAP	mAP@25	Config	Checkpoint
ScanNet val (ResNet50)	47.8	83.6	config	checkpoint
ScanNet val (Swin-B)	50.0	83.6	config	checkpoint
ScanNet test (Swin-B)	47.7	86.2	config	checkpoint

ScanNet Semantic Segmentation

Dataset	mIoU	Config	Checkpoint
ScanNet val (ResNet50)	73.3	config	checkpoint
ScanNet val (Swin-B)	77.8	config	checkpoint
ScanNet test (Swin-B)	74.4	config	checkpoint

Joint 2D-3D on ScanNet and COCO

Model	mAP (ScanNet)	mAP25 (ScanNet)	mAP (COCO)	Config	Checkpoint
ODIN	49.1	83.1	41.2	config	checkpoint

ScanNet200 Instance Segmentation

Dataset	mAP	mAP@25	Config	Checkpoint
ScanNet200 val (ResNet50)	25.6	36.9	config	checkpoint
ScanNet200 val (Swin-B)	31.5	45.3	config	checkpoint
ScanNet200 test (Swin-B)	27.2	39.4	config	checkpoint

ScanNet200 Semantic Segmentation

Dataset	mIoU	Config	Checkpoint
ScanNet200 val (ResNet50)	35.8	config	checkpoint
ScanNet200 val (Swin-B)	40.5	config	checkpoint
ScanNet test (Swin-B)	36.8	config	checkpoint

AI2THOR Semantic and Instance Segmentation

Dataset	mAP	mAP@25	mIoU	Config	Checkpoint
AI2THOR val (ResNet)	63.8	80.2	71.5	config	checkpoint
AI2RHOR val (Swin)	64.3	78.6	71.4	config	checkpoint

Matterport3D Instance Segmentation

Dataset	mAP	mAP@25	Config	Checkpoint
Matterport3D val (ResNet)	11.5	27.6	config	checkpoint
Matterport val (Swin)	14.5	36.8	config	checkpoint

Matterport3D Semantic Segmentation

Dataset	mIoU	mAcc	Config	Checkpoint
Matterport3D val (ResNet)	22.4	28.5	config	checkpoint
Matterport3D val (Swin)	28.6	38.2	config	checkpoint

S3DIS Instance Segmentation

Dataset	mAP	mAP@25	Config	Checkpoint
S3DIS Area5 (ResNet50-Scratch)	36.3	61.2	config	checkpoint
S3DIS Area5 (ResNet50-Fine-Tuned)	44.7	67.5	config	checkpoint
S3DIS Area5 (Swin-B)	43.0	70.0	config	checkpoint

S3DIS Semantic Segmentation

Dataset	mIoU	Config	Checkpoint
S3DIS (ResNet50)	59.7	config	checkpoint
Swin-B	68.6	config	checkpoint

Training Logs:

Please find training logs for all models here

Citing ODIN

If you find ODIN useful in your research, please consider citing:

@misc{jain2024odin,
      title={ODIN: A Single Model for 2D and 3D Perception}, 
      author={Ayush Jain and Pushkal Katara and Nikolaos Gkanatsios and Adam W. Harley and Gabriel Sarch and Kriti Aggarwal and Vishrav Chaudhary and Katerina Fragkiadaki},
      year={2024},
      eprint={2401.02416},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

License

Shield:

The majority of ODIN is licensed under a MIT License.

Acknowledgement

Parts of this code were based on the codebase of Mask2Former and Mask3D.