Enabling RAW Image Classification Using Existing RGB Classifiers
Rasmus Munksø
1,⋆
, Mathias Viborg Andersen
1,⋆
, Lau Nørgaard
2
, Andreas Møgelmose
1
and Thomas B. Moeslund
1
1
Visual Analysis and Perception Lab, Aalborg University, Denmark
2
Phase One A/S, Denmark
Keywords:
RAW, RGB, Transfer Learning, RAW Image Dataset, Classification.
Abstract:
Unprocessed RAW data stands out as a highly valuable image format in image editing and computer vision due
to it preserving more details, colors, and a wider dynamic range as captured directly from the camera’s sensor
compared to non-linearly processed RGB images. Despite its advantages, the computer vision community
has largely overlooked RAW files, especially in domains where preserving precise details and accurate colors
are crucial. This work addresses this oversight by leveraging transfer learning techniques. By exploiting
the vast amount of available RGB data, we enhance the usability of a limited RAW image dataset for image
classification. Surprisingly, applying transfer learning from an RGB-trained model to a RAW dataset yields
impressive performance, reducing the dataset size barrier in RAW research. These results are promising,
demonstrating the potential of cross-domain transfer learning between RAW and RGB data and opening doors
for further exploration in this area of research.
1 INTRODUCTION
Historically, the use of RAW images has mainly been
exploited by photographers wishing to post-process
their captures without losing quality before convert-
ing to RGB images. Despite the advantages to pho-
tographers, RAW files have largely been overlooked
in the computer vision community. RAW Bayer sen-
sor images are minimally processed and thus pre-
serve all the details, colors and dynamic range that
is captured by the camera’s sensor (Yuan and Sun,
2011). The image signal processing (ISP) pipeline
responsible for converting the RAW image to RGB
involves non-linear operations, such as linearization
of the sensor output, white balancing, tone mapping
and gamma correction, as displayed in Figure 1, while
also typically reducing the bit depth of each chan-
nel (Salih et al., 2012; Can and Brown, 2019). This
ISP is designed for human consumption, rather than
capturing accurate physical descriptions of the scene
(Nguyen and Brown, 2017; Wei et al., 2021; Zhang
et al., 2021; Nam et al., 2022). Collectively, the ISP
results in the RGB image becoming a representation
that inevitably loses some of the original capture in-
formation and loses a linear relationship to the physi-
These authors contributed equally
cal brightness of the scene (Salih et al., 2012; Can and
Brown, 2019; Wei et al., 2021).
RAW
image
Linearization,
Demosaicing
RGB
image
White
balancing,
Tone
mapping,
Gamma
correction
Figure 1: Simplified overview of the processes involved in
RAW to RGB image conversion (Kantas et al., 2023).
Given these considerations, it seems glaringly
counter-intuitive that information is discarded with-
out careful consideration. Photographers benefit most
from RAW image post-processing in challenging sce-
narios like extreme exposures and low- or high-
contrasts, where RAW images retain more details
about shadows and highlights, allowing editors to re-
cover otherwise lost information in the RGB images.
One might hypothesize that this loss of significant
data is not limited to these scenarios, but occurs in ev-
ery RGB image at different degrees of severity. Con-
sequently, it would seem logical to leverage the addi-
tional information in the RAW format in the context
Munksø, R., Andersen, M., Nørgaard, L., Møgelmose, A. and Moeslund, T.
Enabling RAW Image Classification Using Existing RGB Classifiers.
DOI: 10.5220/0012363200003660
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 2: VISAPP, pages
123-129
ISBN: 978-989-758-679-8; ISSN: 2184-4321
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
123
of retaining the highest degree of quality in the image
data.
RGB is the predominately used image color rep-
resentation in computer vision. A significant factor
is that RGB images have benefited from advance-
ments in image processing techniques applied to RGB
images and from wide-spread integration within im-
age processing and deep learning workflows and li-
braries. Another factor is due to the limited avail-
ability of RAW image datasets compared to the pres-
ence of large-scale annotated RGB datasets, such as
ImageNet (Deng et al., 2009). Furthermore, RAW
image data suffers from a lack of standardization, as
the RAW formats vary between camera manufacturers
and models, whereas the RGB format follows a well-
defined standard. Perhaps less important is the fact
that RGB images are optimized for human viewing
and thus more intuitive to work with than RAW im-
ages. However, it is relevant to explore whether RGB
images are the optimal image format for all computer
vision tasks, especially in domains where preserving
fine-grained details and accurate color information is
critical. This work aims to address some of these
challenges with using RAW images, by employing
transfer learning to draw on the abundance of avail-
able RGB data, in order to better utilize a small-scale
RAW image dataset for the task of image classifica-
tion. Models trained from scratch are compared to
transfer-learned models to verify the change in per-
formance.
Contribution. The main contributions are:
We introduce an annotated and publicly available
dataset consisting of pairwise RAW images and
their corresponding RGB counterparts, that can be
adapted for a broad range of applications.
We demonstrate that applying transfer-learning
from an RGB-trained model on a small-scale
RAW dataset results in surprisingly good perfor-
mance, while lowering the dataset size barrier for
beginning research into using RAW images.
2 RELATED WORKS
In order to work with RAW images effectively, having
access to an appropriate RAW image dataset is neces-
sary. A thorough search of the existing publicly avail-
able RAW image datasets was conducted and seven
datasets were found (Bychkovsky et al., 2011; Omid-
Zohoor et al., 2014; Dang-Nguyen et al., 2015; Chen
et al., 2018; Zhang et al., 2019; Liang et al., 2020;
Morawski et al., 2022; Kantas et al., 2023). However,
all of these datasets are limited in size.
Currently, to the best of the authors’ knowledge,
there is no open-source, large-scale RAW Bayer im-
age dataset, the lack of which severely hinders re-
search into using RAW images. Therefore, it is es-
sential to acquire more data in order to validate find-
ings and expand RAW image usage. Some studies
have attempted to address this shortage by so-called
RAW image reconstruction, converting existing large-
scale RGB datasets into RAW counterparts. (Nguyen
and Brown, 2017) propose storing RGB-RAW map-
ping parameters in JPEG metadata for future RAW
reconstruction. (Brooks et al., 2019) reverse the ISP
process using camera information like color correc-
tion matrices and digital gains. Other lines of research
employ learning-based techniques to synthesize RAW
data from sRGB images (Liu et al., 2020; Punnap-
purath and Brown, 2020; Waqas Zamir et al., 2020;
Wei et al., 2021). However, these methods rely on the
lossy in-camera ISP pipeline, resulting in slight inac-
curacies compared to the originals. (Xing et al., 2021)
have explored replacing existing ISP pipelines with
invertible ISPs, achieving nearly perfect RAW image
reconstruction. Although this approach is valid, it
does not allow for research into RAW images until
more RAW image data has been gathered and pro-
cessed by their invertible ISP.
Recent research has explored the use of RAW im-
ages in specific vision tasks. (Zhang et al., 2019) fo-
cuses on improving computational zoom using RAW
images. (Chen et al., 2018) aims to train models for
low-light image processing from RAW to final output,
while (Liang et al., 2020) investigates RAW images
for deblurring. These studies collectively demonstrate
that RAW images can outperform RGB in tasks rely-
ing on linear scene radiance properties. Moreover, ob-
ject detection with RAW images has been studied and
shows promise, either by direct RAW image imple-
mentation or by applying learnable non-linear func-
tions as extensions to neural networks (Li et al., 2022;
Morawski et al., 2022; Zhang et al., 2022; Ljung-
bergh et al., 2023). (Kantas et al., 2023) recently ex-
plored image classification using RAW images, show-
ing similar performance for RAW and RGB images,
while demonstrating a significant speed-up for RAW
images by bypassing the ISP pipeline. The tests in
(Kantas et al., 2023) were conducted on a custom,
minimally processed RAW image dataset. Results
were demonstrated by training a ResNet-50 model
from scratch on different representations of RAW im-
age data. These results were compared to an identical
model trained on a corresponding dataset of 8- and
16-bit RGB images.
To the best of the authors’ knowledge, no cur-
rent research has looked into combining the RAW and
VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications
124
RGB domains by using transfer learning. Using a
smaller RAW image dataset together with an RGB
to RAW transfer-learning approach to perform a vi-
sion task, it can be determined whether the dataset
size barrier in RAW image research can be mitigated.
And so, this paper will focus on investigating the fea-
sibility of that approach and whether transfer-learning
from RGB models can improve performance in RAW
image classification.
3 METHODS
In order to investigate the potential of the RAW im-
age format within image classification it would seem
logical to classify data only using the RAW image for-
mat on models trained from scratch. However, given
the scarcity of large-scale RAW image datasets, we
hypothesize that the classification performance on a
small dataset can be improved by applying RGB-to-
RAW transfer learning.
We investigate whether a small-scale RAW dataset
can be trained for image classification while utilizing
transfer learning from an RGB-trained model. There-
fore, some considerations of how to practically use
and implement RAW images into a computer vision
workflow are in order.
3.1 Using RAW Data
The first challenge associated with using RAW im-
ages is the different file formats that are output from
the camera. Very few if any of these are directly sup-
ported within commonly used image processing li-
braries, such as OpenCV and Pillow, or deep learning
libraries, such as TensorFlow or PyTorch. However,
specific image processing libraries exist that do sup-
port RAW image formats. In this work, the Python
wrapper for LibRaw called RawPy (Riechert, 2014)
is used to extract the RAW image data as a simple 2-
dimensional data matrix without any processing ap-
plied.
The next challenge is storing the RAW image data.
While possible, storing a RAW image in a lossless
standard RGB format such as PNG or TIFF would
be inefficient. Additionally, this work uses a differ-
ent representation of RAW data obtained by packing
the original RAW data, where each RAW image is re-
arranged into the four color channels present in the
Bayer pattern. As this results in a 4-channel RAW
image representation, standard image formats seem
unsuitable. For this reason, both types of RAW im-
ages are stored as NumPy array files in this work. The
practice of saving RAW image data in NumPy files
has the added benefit of making it straightforward to
efficiently load data when standardized data loaders in
libraries such as TensorFlow or PyTorch do not sup-
port the chosen image file format. The approach taken
in this work is to load each NumPy file and concate-
nate the data to one large array, from which a dataset
loader can be created in both PyTorch and Tensor-
Flow.
3.2 Dataset
To use the full-resolution RAW images from the
datasets (research on datasets can be found in Sec-
tion 2) as inputs to a classifier, downscaling would be
required, which would artificially process the origi-
nal image, resulting in a lower-quality representation.
This would be counter-intuitive for this work since
one of the primary reasons for using RAW images is
that they contain unaltered capture information. For
these reasons, a more relevant dataset for investigat-
ing the hypothesis consists of small, unaltered RAW
images. Since such a dataset is not publicly available,
we create a custom dataset based on classes within the
PASCALRAW dataset (Omid-Zohoor et al., 2014).
The PASCALRAW dataset has the added benefit of
being captured using a Nikon D3200 DSLR camera
and therefore also being stored in the same Nikon
Raw Image file format (NEF), removing the need for
considerations for differences in capture data and im-
age formats.
Figure 2: Grid of crops overlapping the semantic mask with
at least 50 % of its pixels.
Our dataset is created by extracting 448x448-
https://www.kaggle.com/datasets/mathiasviborg/pasc
alraw-derived-object-cropped-dataset
Enabling RAW Image Classification Using Existing RGB Classifiers
125
sized cropped images of people, bicycles, cars and
backpacks that are present in each full-resolution
RAW- and corresponding RGB images within the
PASCALRAW dataset. The dataset-generating pro-
cess utilizes YOLOv5 (Jocher et al., 2022) to apply
instance segmentation to extract masks of each of the
desired four classes that are present in each RGB im-
age. By then placing a grid of cropped images of
size 448x448 over each object’s entire bounding box,
each crop overlapping itself with a stride of 224 pix-
els along its x- and y-direction, see Figure 2, and only
storing those crops that overlap the semantic mask
with at least 50 % of its pixels, data samples from
both the RGB- and corresponding RAW image are ex-
tracted, see Figure 3.
Figure 3: Overview of the first four extracted samples from
the object presented in Figure 2. Shown in Original RGB
and Original RAW (RAW displayed using Matplotlib).
As one might expect, the majority of these sam-
ples do not contain the entire object. However, in this
dataset, the aim is to store samples showing approxi-
mately a quarter of the object at minimum as a trade-
off between maintaining adequate semantic content
within each sample and extracting a sufficient number
of samples for the dataset. This is achieved by limit-
ing the size of the bounding boxes to be considered
and manual inspection of the gathered samples.
W/2
H/2
W
H
Packed RAW
Original RAW
Original RGB Resized RGB
448 x 448 x 1 224 x 224 x 4 448 x 448 x 3 224 x 224 x 3
W
H
W/2
H/2
Figure 4: Overview of the different image representations
in the dataset as well as their respective dimensions.
After packing the RAW image samples the height
and width of the samples are halved, which by design
corresponds to the input resolution of the chosen clas-
sifier, ResNet-50 (He et al., 2016). The RGB samples
are simply resized and all the four types of samples
are stored in a dataset consisting of data formats as
shown in Figure 4.
Figure 5: The 4 classes in the dataset. From left: Backpack,
Bicycle, Car and Person. Shown in Original RGB and Orig-
inal RAW (RAW displayed using Matplotlib).
The distribution of the classes within the dataset
can be found in Table 1 and examples of samples from
each class are shown in Figure 5.
Table 1: Sample distribution of the dataset.
Backpack Bicycle Car Person
3569 5142 4856 4864
3.3 Implementation
ResNet-50 pretrained on ImageNet (Deng et al.,
2009) is chosen as the CNN model as it is consid-
ered a well-performing general purpose classifier and
is trained on RGB data. The transfer-learning fine-
tuning is run on an Nvidia A40 GPU. The hyperpa-
rameters of the networks, as found in Table 2, are de-
termined by a hyperparameter sweep conducted using
the Weights & Biases Bayesian Hyperparameter Opti-
mization tool (Biewald, 2020). The hyperparameters
were then tested to ensure performance and stability
during training.
Table 2: Overview of the hyperparameters used for transfer
learning vs. learning from scratch.
RAW
PACKED
RAW
RGB
RESIZED
RGB
Trained via. Transfer Learning
Optimizer Adam Adam Adam Adam
Learning
Rate
0.0427 0.0489 0.0628 0.0409
Batch
Size
8 128 32 32
Trained from Scratch
Optimizer Adam Adam Adam Adam
Learning
Rate
0.0428 0.0715 0.0629 0.0319
Batch
Size
8 8 32 8
The chosen architecture can be seen on Figure 6
and trainable parameters on Table 3.
VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications
126
Pre-Trained
ResNet-50
Input
224x224x3
7x7,64
Conv1
(112x112 )
7x7,64
Conv1
(224x224)
1x1,64
3x3,64
1x1,256
Conv2
(56x56)
Conv2
(112x112 )
1x1,64
3x3,64
1x1,256
Layer Name
(Output size)
Layer Name
(Output size)
1x1,128
3x3,128
1x1,512
Conv3
(28x28)
Conv3
(56x56)
1x1,128
3x3,128
1x1,512
Pre-Trained
ResNet-50
Input
448x448x3
1x1,256
3x3,256
1x1,1024
Conv4
(14x14)
Conv4
(28x28)
1x1,256
3x3,256
1x1,1024
1x1,512
3x3,512
1x1,2048
Conv5
(7x7)
Conv5
(14x14)
1x1,512
3x3,512
1x1,2048
x 3
x 4
x 6
x 3
x 3
x 4
x 6
x 3
Global
Average
Pooling
Global
Average
Pooling
FC 4 FC 4
1x1,3 1x1,3 1x1,3 1x1,3
Packed RAW
Image
Input
224x224x1
Resized RGB
Image
Input
224x224x3
Original RAW
Image
Input
448x448x1
Original RGB
Image
Input
448x448x3
Figure 6: The network architectures for each image repre-
sentation. Note that models are trained separately for each
image representation and that inputs are not fused (sim-
ply visualized as such for compactness). Visualized with
ResNet-50 architecture (He et al., 2016) shown with the
residual units, the size of the filters and the outputs of each
convolutional layer. Key: The notation k × k, n in the con-
volutional layer block denotes a filter of size k and n chan-
nels. FC 4 denotes the fully connected layer with 4 neurons
representing the 4 classes. The number to the right of the
convolutional layer block represents the repetition of each
unit. Red colored blocks denote frozen layers that are not
updated during training when using transfer learning, while
green colored blocks denote layers with learnable parame-
ters.
It can be seen from Figure 6 that the only addi-
tion to the architecture apart from the RGB-trained
ResNet-50 is a convolutional layer with a 1x1 kernel
and 3 filters, resulting in the dimensions (224x224x3)
Table 3: Number of trainable parameters for transfer learn-
ing vs. learning from scratch.
RAW
PACKED
RAW
RGB
RESIZED
RGB
Trained via. Transfer Learning
8.202 K 8.202 K 8.208 K 8.208 K
Trained from Scratch
23.52 M 23.52 M 23.54 M 23.54 M
for Packed RAW and Resized RGB, which is the de-
fault input dimension for the pre-trained model. Note
that the output size from each convolutional layer is
different for packed and resized as opposed to origi-
nal input images, while the learned weights in the pre-
trained convolutional layers are identical. It has been
experimentally observed that the input dimensions of
the Original RAW and Original RGB perform well
in this architecture, even though the input dimension
to the pre-trained ResNet-50 model is different from
what it was trained on. As it performed better than
introducing a down-sampling convolutional layer, the
architecture was kept as is.
4 RESULTS
Experiments are conducted to test whether transfer
learning from RGB trained models can be used on
RAW images. The impact of transfer learning on im-
proving validation accuracies has been thoroughly in-
vestigated and compared to training from scratch. The
networks are trained 10 times each on the selected
data types (Original-RAW, Packed-RAW, RGB, and
Resized-RGB) using the previously stated hyperpa-
rameters from Tabel 2. The average of the 10 top-
1 accuracies is stored as the final top-1 classification
accuracy, see Table 4.
Table 4: Mean top-1 classification validation accuracies as
measured from the models with lowest validation loss.
RAW
PACKED
RAW
RGB
RESIZED
RGB
Trained via. Transfer Learning
Mean Top-1
Accuracy
96.04 % 94.27 % 97.25 % 96.79 %
Standard
Deviation
0.03 0.12 0.09 0.05
Trained from Scratch
Mean Top-1
Accuracy
85.64 % 82.92 % 86.54 % 87.49 %
Standard
Deviation
0.139 0.013 0.027 0.021
Enabling RAW Image Classification Using Existing RGB Classifiers
127
5 DISCUSSION AND FUTURE
WORKS
The classification accuracy results found in Table 4
demonstrate that cross-domain transfer learning from
RGB-trained models has a significant improvement
on RAW image classification, with both RAW im-
age representations gaining more than 10 percent-
age points. The difference in performance for the
different RAW representations is in coherence with
the found results in (Kantas et al., 2023) as Origi-
nal RAW performs better than Packed RAW, although
by a higher margin. This raises the question of
whether the use of Packed RAW is warranted on its
own. However, in previous works, RAW images have
been represented by combining Original RAW and
Packed RAW by a Bidirectional Cross-Modal (BCA)
approach (Liang et al., 2020)RAWInstinct. This rep-
resentation is not investigated in this work, as the pri-
mary research objective is not to outperform RGB in
classification tasks. However, it is worth noting that
this alternative representation might have improved
image classification performance compared to RGB,
as demonstrated in prior work (Kantas et al., 2023).
Despite the increased performance from transfer
learning, the RAW image data performs similarly
to but not surpassing the RGB image data. How-
ever, the objective is not for RAW to outperform the
RGB image classification results. Instead, these re-
sults demonstrate that utilizing RGB-trained models
on RAW image data can increase the performance of
a small-scale RAW image dataset compared to the
performance when training from scratch. This ap-
proach effectively lowers the dataset barrier, mak-
ing RAW image research more accessible and time-
efficient. Therefore, these findings are promising,
showcasing the effectiveness of cross-domain transfer
learning between RAW and RGB data and suggesting
potential avenues for further exploration in this area.
In this work, we generate a dataset by extract-
ing samples of backpacks, people, bicycles and cars
from an existing RAW image object detection dataset
(Omid-Zohoor et al., 2014). Due to the nature of ex-
tracting small-sized images from objects within large-
scale captures, few of the small-sized crops have full-
sized objects within their frame, instead showing a
cropped-out portion of it. This adds a built-in way
of data augmentation, however, it might not be ideal
for the purpose of investigating fine details in RAW
images compared to RGB images. Furthermore, only
a minority of the images are shot in challenging sce-
narios in the original PASCALRAW dataset and an
argument could therefore be made that the created
dataset does not fully explore the advantages of RAW
images. This could explain the lack of observable
improvement in performance when classifying RAW
images compared to RGB in the results presented in
this work. As the results found in this paper are only
demonstrated on this one dataset, further research on
more intricate datasets is necessary to generalize these
findings effectively and demonstrate which types of
images may benefit the most from RAW image clas-
sification.
In future works, researchers may endeavor to con-
struct a dataset specifically designed to highlight the
advantages of RAW images. Such a dataset could
include images captured in challenging lighting sce-
narios, resulting in severely under- or overexposed
images, images with low contrast, among other sce-
narios. Guided by the premise that finer details are
preserved in the RAW image format compared to the
RGB format, leveraging the advantages of RAW im-
ages in classifying such challenging scenarios could
yield substantial benefits. Therefore it would be inter-
esting to research whether the fine details in the RAW
image format could result in better classification ac-
curacies through transfer learning compared to those
of corresponding RGB images.
6 CONCLUSION
This work shows that using RGB-trained models for
RAW image classification is can be effective. The re-
sults indicate that transferring knowledge from RGB
models significantly improves accuracy. By tapping
into this existing knowledge cross-domain transfer
learning is demonstrated to not only enhance accu-
racy for image classification but also potentially make
future research into the use of RAW images more ac-
cessible.
ACKNOWLEDGEMENTS
The work was financially supported by the
AI:Denmark project funded by the Danish Industry
Foundation.
REFERENCES
Biewald, L. (2020). Experiment tracking with weights and
biases. Software available from wandb.com.
Brooks, T., Mildenhall, B., Xue, T., Chen, J., Sharlet, D.,
and T Barron, J. (2019). Unprocessing images for
learned raw denoising. In Conference on Computer
Vision and Pattern Recognition.
VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications
128
Bychkovsky, V., Paris, S., and E. Chan, F. D. (2011).
Learning photographic global tonal adjustment with a
database of input / output image pairs. In Conference
on Computer Vision and Pattern Recognition.
Can, H. and Brown, M. (2019). Beyond raw-rgb and srgb:
Advocating access to a colorimetric image state. In
Color and Imaging Conference.
Chen, C., Chen, Q., Xiu, J., and Koltun, V. (2018). Learning
to see in the dark. In Conference on Computer Vision
and Pattern Recognition.
Dang-Nguyen, D.-T., Pasquini, C., Conotter, V., and Boato,
G. (2015). Raise - a raw images dataset for digital
image forensics. In ACM Multimedia Systems Confer-
ence.
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-
Fei, L. (2009). Imagenet: A large-scale hierarchical
image database. In Conference on Computer Vision
and Pattern Recognition.
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
ual learning for image recognition. In Conference on
Computer Vision and Pattern Recognition.
Jocher, G., Chaurasia, A., Stoken, A., Borovec, J.,
NanoCode012, Kwon, Y., Michael, K., TaoXie, Fang,
J., imyhxy, Lorna, Yifu, Z., Wong, C., V, A., Montes,
D., Wang, Z., Fati, C., Nadar, J., Laughing, UnglvK-
itDe, Sonck, V., tkianai, yxNONG, Skalski, P., Hogan,
A., Nair, D., Strobel, M., and Jain, M. (2022). ultralyt-
ics/yolov5: v7.0 - YOLOv5 SOTA Realtime Instance
Segmentation.
Kantas, C., Antoniussen, B., Andersen, M. V., Munksø, R.,
Kotnala, S., Jensen, S. B., Møgelmose, A., Nørgaard,
L., and Moeslund, T. B. (2023). Raw instinct: Trust
your classifiers and skip the conversion. In Conference
on Pattern Recognition and Artificial Intelligence.
Li, Z., Lu, M., Zhang, X., Feng, X., Asif, M. S., and Ma, Z.
(2022). Efficient visual computing with camera raw
snapshots. In Conference on Computer Vision and
Pattern Recognition.
Liang, C., Chen, Y., Liu, Y., and Hsu, W. H. (2020). Raw
image deblurring. In IEEE Transactions on Multime-
dia, volume 24, pages 61–72.
Liu, Y.-L., Lai, W.-S., Chen, Y.-S., Kao, Y.-L., Yang, M.-H.,
Chuang, Y.-Y., and Huang, J.-B. (2020). Single-image
hdr reconstruction by learning to reverse the camera
pipeline. In Conference on Computer Vision and Pat-
tern Recognition.
Ljungbergh, W., Johnander, J., Petersson, C., and Felsberg,
M. (2023). Raw or cooked? object detection on raw
images. In Scandinavian Conference on Image Anal-
ysis.
Morawski, I., Chen, Y.-A., Lin, Y.-S., Dangi, S., He, K.,
and Hsu, W. H. (2022). Genisp: Neural isp for low-
light machine cognition. In Conference on Computer
Vision and Pattern Recognition.
Nam, S., Punnappurath, A., Brubaker, M. A., and Brown,
M. S. (2022). Learning srgb-to-raw-rgb de-rendering
with content-aware metadata. In Conference on Com-
puter Vision and Pattern Recognition.
Nguyen, R. and Brown, M. (2017). Raw image reconstruc-
tion using a self-contained srgb–jpeg image with small
memory overhead. In International Journal of Com-
puter Vision, volume 126(6), pages 637–650.
Omid-Zohoor, A., Young, C., and D. Ta, B. M. (2014). Pas-
calraw: Raw image database for object detection in
searchworks catalog. In Stanford Digital Repository.
Punnappurath, A. and Brown, M. S. (2020). Learning raw
image reconstruction-aware deep image compressors.
In IEEE Transactions on Pattern Analysis and Ma-
chine Intelligence, volume 42(4), pages 1013–1019.
Riechert, M. (2014). rawpy: Raw image processing for
python, a wrapper for libraw. Software available from
https://github.com/letmaik/rawpy.
Salih, Y., bt. Md-Esa, W., Malik, A. S., and Saad, N. (2012).
Tone mapping of hdr images: A review. In Interna-
tional Conference on Intelligent and Advanced Sys-
tems.
Waqas Zamir, S., Arora, A., Khan, S., Hayat, M., Shah-
baz Khan, F., Yang, M.-H., and Shao, L. (2020). Cy-
cleisp: Real image restoration via improved data syn-
thesis. In Conference on Computer Vision and Pattern
Recognition.
Wei, Z., Xiangyu, Z., Hongyu, W. S. G., and Xin, L.
(2021). Raw bayer pattern image synthesis for com-
puter vision-oriented image signal processing pipeline
design. In Conference on Computer Vision and Pat-
tern Recognition.
Xing, Y., Qian, Z., and Chen, Q. (2021). Invertible image
signal processing. In Conference on Computer Vision
and Pattern Recognition.
Yuan, L. and Sun, J. (2011). High quality image reconstruc-
tion from raw and jpeg image pair. In International
Conference on Computer Vision.
Zhang, X., Chen, Q., Ng, R., and Koltun, V. (2019). Zoom
to learn, learn to zoom. In Conference on Computer
Vision and Pattern Recognition.
Zhang, X., Zhang, L., and Lou, X. (2022). A raw image-
based end-to-end object detection accelerator using
hog features. In IEEE Transactions on Circuits and
Systems I: Regular Papers, volume 69(1), pages 322–
333.
Zhang, Z., Wang, H., Liu, M., Wang, R., Zhang, J., and Zuo,
W. (2021). Learning raw-to-srgb mappings with inac-
curately aligned supervision. In International Confer-
ence on Computer Vision.
Enabling RAW Image Classification Using Existing RGB Classifiers
129