Official PyTorch implementation of N-ImageNet: Towards Robust, Fine-Grained Object Recognition with Event Cameras (ICCV 2021) [Paper] [Video].
In this repository, we provide instructions for downloading N-ImageNet along with the implementation of the baseline models presented in the paper. If you have any questions regarding the dataset or the baseline implementations, please leave an issue or contact [email protected].
🌟 Update 1 🌟 Answering questionnaires are no longer necessary to download N-ImageNet! Please refer to the instructions below to download the dataset.
🌟 Update 2 🌟 Check out the public benchmark on object recognition and robust classification available at the following link. Feel free to upload new results to the benchmark!
🌟 Update 3 🌟 We have newly released mini N-ImageNet 👶! The dataset contains 100 classes, which is 1/10 of the original N-ImageNet. We expect the dataset to enable quick and light-weight evaluation of new event-based object recognition methods. To download the dataset, please refer to the instructions stated here. To download the pretrained models, check here.
🌟 Update 4 🌟 We finally fixed the download issues with full N-ImageNet! Now N-ImageNet can be easily downloaded from HuggingFace Datasets 🤗. Detailed instructions are provided below.
We maintain a publicly available benchmark for N-ImageNet at the following link. Feel free to upload new results to the benchmark!
Currently we have three benchmarks available.
- Classification on entire N-ImageNet: Here we report the classfication accuracy measured on the original validation split from N-ImageNet.
- Classification on N-ImageNet variants: Here we report the average classification accuracy measured on the nine N-ImageNet variants. The N-ImageNet variants are recorded on various camera trajectory and lighting, and the details are further specified here.
- Classification on mini N-ImageNet: Here we report the classification accuracy on the mini original validation split which contains 100 classes. The models are assumed to be trained on the mini N-ImageNet train split which also contains the same number of classes.
N-ImageNet can be downloaded from Huggingface Datasets: follow the link here. Please refer to the following instructions to learn about how the dataset is organized. If you have any additional questions regarding the dataset, drop an email to [email protected].
You can directly download mini N-ImageNet from here. For gaining access to the other mini validation splits from the N-ImageNet variants, please refer to the following link.
The codebase is tested on a Ubuntu 18.04 machine with CUDA 10.1. However, it may work with other configurations as well. First, create and activate a conda environment with the following command.
conda env create -f environment.yml
conda activate e2t
In addition, you must install pytorch_scatter. Follow the instructions provided in the pytorch_scatter github repo. You need to install the version for torch 1.7.1 and CUDA 10.1.
Before you move on to the next step, please download N-ImageNet along with train_list.txt and val_list.txt from here. Once you download N-ImageNet, you will spot a structure as follows. Note: If you are using mini N-ImageNet, after downloading you will need to re-structure the directory as below, namely move all the validation data below extracted_val and all the training data below extracted_train.
N_Imagenet
├── train_list.txt
├── val_list.txt
├── extracted_train (train split)
│ ├── nXXXXXXXX (label)
│ │ ├── XXXXX.npz (event data)
│ │ │
│ │ ⋮
│ │ │
│ │ └── YYYYY.npz (event data)
└── extracted_val (val split)
└── nXXXXXXXX (label)
├── XXXXX.npz (event data)
│
⋮
│
└── YYYYY.npz (event data)
The N-ImageNet variants file (which would be saved as N_Imagenet_cam once downloaded) will have a similar file structure, except that it only contains validation files.
The following instruction is based on N-ImageNet, but one can follow a similar step to test with N-ImageNet variants.
First, modify train_list.txt and val_list.txt such that it matches the directory structure of the downloaded data.
To illustrate, if you open train_list.txt you will see the following
/home/jhkim/Datasets/N_Imagenet/extracted_train/n01440764/n01440764_10026.npz
⋮
/home/jhkim/Datasets/N_Imagenet/extracted_train/n15075141/n15075141_999.npz
Modify each path within the .txt file so that it accords with the directory in which N-ImageNet is downloaded.
For example, if N-ImageNet is located in /home/user/assets/Datasets/, modify train.txt as follows.
/home/user/assets/Datasets/N_Imagenet/extracted_train/n01440764/n01440764_10026.npz
⋮
/home/user/assets/Datasets/N_Imagenet/extracted_train/n15075141/n15075141_999.npz
In addition, download the Imagenet/ folder from here which contains the development kit text files needed to run the code below.
Once this is done, create a Datasets/ directory within real_cnn_model, and create a symbolic link within Datasets.
To illustrate, using the directory structure of the previous example, deploy the following command.
cd PATH_TO_REPOSITORY/real_cnn_model
mkdir Datasets; cd Datasets
ln -sf /home/user/assets/Datasets/Imagenet/ ./
ln -sf /home/user/assets/Datasets/N_Imagenet/ ./
ln -sf /home/user/assets/Datasets/N_Imagenet_cam/ ./ (If you have also downloaded the variants)
Congratulations! Now you can start training/testing models on N-ImageNet.
You can train a model based on the binary event image representation with the following command.
export PYTHONPATH=PATH_TO_REPOSITORY:$PYTHONPATH
cd PATH_TO_REPOSITORY/real_cnn_model
python main.py --config configs/imagenet/cnn_adam_acc_two_channel_big_kernel_random_idx.ini
For the examples below, we assume the PYTHONPATH environment variable is set as above.
Also, you can change minor details within the config before training by using the --override flag.
For example, if you want to change the batch size use the following command.
python main.py --config configs/imagenet/cnn_adam_acc_two_channel_big_kernel_random_idx.ini --override 'batch_size=8'
In addition, if you want to train a model using a different event representation, for example timestamp image, use the following command:
python main.py --config configs/imagenet/cnn_adam_acc_two_channel_big_kernel_random_idx.ini --override 'loader_type=timestamp_image'
For training models on the mini N-ImageNet dataset, use the following command. Note that we provide the mini-counterparts for all the configs as configs with additional _mini prefixes attached in the configs/ folder.
export PYTHONPATH=PATH_TO_REPOSITORY:$PYTHONPATH
cd PATH_TO_REPOSITORY/real_cnn_model
python main.py --config configs/imagenet/cnn_adam_acc_two_channel_big_kernel_random_idx_mini.ini
Similar to the example above, one can change the event representation with the override flag. For example, to train using DiST, use the following command:
python main.py --config configs/imagenet/cnn_adam_acc_two_channel_big_kernel_random_idx_mini.ini --override 'loader_type=dist'
Suppose you have a pretrained model saved in PATH_TO_REPOSITORY/real_cnn_model/experiments/best.tar.
You can evaluate the performance of this model on the N-ImageNet validation split by using the following command.
python main.py --config configs/imagenet/cnn_adam_acc_two_channel_big_kernel_random_idx.ini --override 'load_model=PATH_TO_REPOSITORY/real_cnn_model/experiments/best.tar'
For a new representation (e.g. timestamp image), one should also change the loader_type as follows:
python main.py --config configs/imagenet/cnn_adam_acc_two_channel_big_kernel_random_idx.ini --override 'load_model=PATH_TO_REPOSITORY/real_cnn_model/experiments/best.tar,loader_type=timestamp_image'
Similar to the full N-ImageNet dataset, suppose you have a pretrained model saved in PATH_TO_REPOSITORY/real_cnn_model_mini/experiments/best.tar.
You can evaluate the performance of this model on the mini N-ImageNet validation split by using the following command.
python main.py --config configs/imagenet/cnn_adam_acc_two_channel_big_kernel_random_idx_mini.ini --override 'load_model=PATH_TO_REPOSITORY/real_cnn_model_mini/experiments/best.tar'
For a new representation (e.g. timestamp image), one should also change the loader_type as follows:
python main.py --config configs/imagenet/cnn_adam_acc_two_channel_big_kernel_random_idx_mini.ini --override 'load_model=PATH_TO_REPOSITORY/real_cnn_model_mini/experiments/best.tar,loader_type=timestamp_image'
The naming of event representations used in the codebase is different from that of the original paper. Please use the following table to convert event representations used in the paper to event representations used in the codebase. Note that when specifying the representation names in the config .ini, you can also use the alias names shown in parentheses.
| Paper | Codebase |
|---|---|
| DiST | reshape_then_acc_adj_sort (alias dist, DiST) |
| Binary Event Image | reshape_then_acc_flat_pol (alias binary_event_image) |
| Event Image | reshape_then_acc (alias event_image) |
| Timestamp Image | reshape_then_acc_time_pol (alias timestamp_image) |
| Event Histogram | reshape_then_acc_count_pol (alias event_histogram) |
| Sorted Time Surface | reshape_then_acc_sort (alias sorted_time_surface) |
One can download the pretrained models on the N-ImageNet dataset through the following links. Here we contain pretrained models and the configs used to train them.
If you find the dataset or codebase useful, please cite
@InProceedings{Kim_2021_ICCV,
author = {Kim, Junho and Bae, Jaehyeok and Park, Gangin and Zhang, Dongsu and Kim, Young Min},
title = {N-ImageNet: Towards Robust, Fine-Grained Object Recognition With Event Cameras},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month = {October},
year = {2021},
pages = {2146-2156}
}