Implementation of : HIC-YOLOv5: Improved YOLOv5 for Small Object Detection

[aash1999]

This repository contains the code for HIC-YOLOv5, an improved version of YOLOv5 tailored for small object detection. The improvements are based on the paper HIC-YOLOv5: Improved YOLOv5 For Small Object Detection.

HIC-YOLOv5 incorporates Channel Attention Block (CBAM) and Involution modules for enhanced object detection, making it suitable for both CPU and GPU training.

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

📊 Key Changes

• New hyperparameter file for small object detection, hyp.hic-yolov5s.yaml, tailored to the VisDrone Dataset.
• Introduction of ChannelAttention and SpatialAttention modules in common.py to enhance feature representation.
• Implementation of CBAM module, combining channel and spatial attention for richness in feature maps.
• Addition of Involution module, a novel operation to address limitations of convolutions.
• Creation of yolov5s-cbam-involution.yaml architecture with CBAM and Involution integrated into the YOLOv5s model.

🎯 Purpose & Impact

• The PR aims to improve YOLOv5's ability to detect small objects, a common challenge in drone and surveillance applications.
• Attention mechanisms (CBAM) and involution help in capturing better feature representations without significantly increasing computational cost.
• Users can expect improved performance on datasets with small objects without major changes to their existing workflows.

🌟 Summary

"YOLOv5 enhancements with attention mechanisms and involution for boosting small object detection performance." 🛸🔍

[github-actions]

👋 Hello @aash1999, thank you for submitting a YOLOv5 🚀 PR! To allow your work to be integrated as seamlessly as possible, we advise you to:

• ✅ Verify your PR is up-to-date with ultralytics/yolov5 master branch. If your PR is behind you can update your code by clicking the 'Update branch' button or by running git pull and git merge master locally.
• ✅ Verify all YOLOv5 Continuous Integration (CI) checks are passing.
• ✅ Reduce changes to the absolute minimum required for your bug fix or feature addition. "It is not daily increase but daily decrease, hack away the unessential. The closer to the source, the less wastage there is." — Bruce Lee

[Ilyabasharov]

@aash1999 Hello! Thanks for your work! Did you evaluate the results? What about metrics?

[aash1999]

Hi @Ilyabasharov
I implemented as mentioned by the paper : https://arxiv.org/pdf/2309.16393.pdf
You can find the results as mentioned in them. And also I am training for the same to validate the results.

thank you

[glenn-jocher]

Hi @aash1999,

Thank you for your interest in HIC-YOLOv5! As mentioned in the paper, the results and metrics can be found by referring to the research paper at https://arxiv.org/pdf/2309.16393.pdf. You can find the specific details and evaluations there. Additionally, I am currently training and validating the model to further validate the results.

If you have any further questions or need assistance, feel free to ask.

Thank you!

[aash1999]

Hi @glenn-jocher

Thank you for your prompt response. I'm currently encountering some issues with the checks, as they are failing. Could you please provide some guidance on how I can rectify these issues and ensure the checks pass successfully?

Your assistance is greatly appreciated.

Thank you

[Ilyabasharov]

Hi @Ilyabasharov I implemented as mentioned by the paper : https://arxiv.org/pdf/2309.16393.pdf You can find the results as mentioned in them. And also I am training for the same to validate the results.

thank you

did you manage to reproduce the metrics from the article? Im also interested in this results, but it seems according to the article that tph-yolov5 paper, github gives better performance on VisDrone Dataset

[aash1999]

@Ilyabasharov

I am currently training the model with mentioned hyper parameter, but I am not able to run on the batch size that was mentioned due to GPU Ram constraint.

[glenn-jocher]

@aash1999 hi,

Thank you for reaching out. I understand that you are experiencing GPU RAM constraints while trying to run the model with the mentioned batch size. GPU RAM limitations can indeed be a challenge.

To address this issue, you can try the following potential solutions:

1. Reduce the batch size: You can decrease the batch size until it fits within the available GPU RAM. However, please keep in mind that reducing the batch size may affect training performance.

2. Utilize gradient accumulation: Instead of updating the model weights after every batch, you can accumulate gradients over multiple batches before performing a weight update. This allows you to effectively simulate a larger batch size without exceeding the GPU RAM limit.

3. Utilize mixed precision training: By using mixed precision training, you can take advantage of GPU tensor cores and reduce the memory requirement. Tools like Nvidia's Automatic Mixed Precision (AMP) can help streamline this process.

Please note that these are general suggestions, and the optimal solution may vary depending on your specific use case and the resources available to you.

I hope this information helps! If you have any further questions or need additional assistance, please let me know.

Thank you!

[aash1999]

Hi @glenn-jocher, @Ilyabasharov

I ran the model on an A100 for 300 epochs as mentioned in the paper with the same hyperparameters, but only with a batch size of 70. I obtained the following results:

HIC-YOLOv5 (test): 35.16 mAP@0.5, 20.23 mAP@[0.5:0.95]
HIC-YOLOv5 (Val): 44.02 mAP@0.5, 25.82 mAP@[0.5:0.95]

Meanwhile, YOLOv5 (test) achieved 27.57 mAP@0.5 and 14.43 mAP@[0.5:0.95] on the VISDRONE dataset.

Thanks

[aash1999]

@glenn-jocher @Ilyabasharov
are we good for merging ?

[glenn-jocher]

@aash1999 thank you for considering merging the changes. We appreciate your contribution to the YOLOv5 repository. Before merging, we need to ensure that the changes align with the project's guidelines and requirements.

Please provide more details about the changes you made and any relevant information, such as how the changes impact the overall functionality and performance of the model. Once we have a clearer understanding, we can proceed with the review process and determine if the changes are ready for merging.

Thank you again for your contribution. We look forward to reviewing your changes.

[aash1999]

Hi @glenn-jocher,

I'd like to provide you with a comprehensive overview of the code changes I've made in this PR:

1. CBAM and Involution Modules: In this update, I introduced two crucial modules - CBAM and Involution. To incorporate these modules into the YOLOv5 model, I made the following modifications:

   • In models/common.py, I added the necessary code to integrate the CBAM module into the backbone of the model. CBAM relies on two other modules: ChannelAttention and SpatialAttention, which are also implemented in common.py. I've documented these changes to provide clarity.

   • To facilitate the integration of Involution, I included code in the prediction head of the model. This ensures that both CBAM and Involution are seamlessly woven into the architecture.

   These changes have been made while maintaining the existing workflow's performance and functionality.

2. Modification in model/yolo.py: To facilitate the parsing of CBAM and Involution, I made a single change in model/yolo.py. This adjustment ensures a smooth flow and compatibility with the newly added modules.

3. Configuration Files: I added two configuration files - models/yolov5s-cbam-involution.yaml and data/hyps/cbam.hyp.yaml. These files are integral for implementing the architecture as described in the paper.

4. Update in utils/general.py: To address an issue related to the nn.AdaptiveAvgPool2d function, which lacks a backward implementation during GPU training, I made a small change in utils/general.py. Specifically, I included the code torch.use_deterministic_algorithms(False, warn only=True). If there's an alternative solution to tackle this problem, I'm open to making those adjustments.

These code changes have been meticulously designed to enhance the YOLOv5 model by incorporating CBAM and Involution modules, as outlined in the referenced paper.

For Performance refer to : #12264 (comment)

[glenn-jocher]

@aash1999 hi,

Thank you for providing a comprehensive overview of the code changes you made in this PR. I appreciate the effort you put into integrating the CBAM and Involution modules into the YOLOv5 model.

I have carefully reviewed your changes, and they seem well-documented and aligned with the goals of enhancing the model's performance. I also took a look at the performance metrics you shared in the linked comment, and the results look promising.

The modifications you made in models/common.py and model/yolo.py to integrate the CBAM and Involution modules, as well as the addition of the two configuration files, appear to be well thought out and essential for implementing the architecture described in the paper.

I see that you also addressed an issue related to backward implementation during GPU training in utils/general.py by including the code torch.use_deterministic_algorithms(False, warn only=True). If there are any alternative solutions to tackle this problem, it would be beneficial to explore them.

Overall, I think your changes align with our project's objectives and will enhance the performance of YOLOv5. However, before merging, I would appreciate it if you could address any open issues and ensure that all tests and checks pass successfully.

Thank you for your contribution. Keep up the great work!

Best,

[aash1999]

@glenn-jocher
Thank you for your efforts in reviewing. I am encountering an issue with the checks. Even though it passes all the tests, there is one in the pre-commit:

fix end of files.........................................................Passed
trim trailing whitespace.................................................Failed
- hook id: trailing-whitespace
- exit code: 1
- files were modified by this hook

Fixing utils/general.py

check for case conflicts.................................................Passed
check docstring is first.................................................Passed
fix double quoted strings................................................Passed
detect private key.......................................................Passed
Upgrade code.............................................................Passed
Sort imports.............................................................Passed
YAPF formatting..........................................................Failed
- hook id: yapf
- files were modified by this hook
MD formatting............................................................Passed
PEP8.....................................................................Passed
codespell................................................................Passed

it will be helpful if you guide me on how o fix it .

[aash1999]

@glenn-jocher Hi

I did the changes as you mentioned and now its passing all the checks.
its ready for merging.

thanks

[glenn-jocher]

Hi @aash1999,

Thank you for making the necessary changes and addressing the issues with the pre-commit checks. I'm glad to hear that the modifications have passed all the checks and that your changes are now ready for merging.

Your contribution is greatly appreciated. I will review your changes again and proceed with the merging process if everything looks good. Once merged, the enhancements you made to the YOLOv5 model will be available for everyone to benefit from.

Thank you again for your hard work and dedication. Keep up the excellent work!

Best,

[aash1999]

@glenn-jocher hi
Any updates on merging this branch to master ?
thanks

[aash1999]

@glenn-jocher
Hi any updates on merging it to master brach as this PR was opened 2 weeks before ?

[glenn-jocher]

@aash1999 hi,

Thank you for your patience. The merging process for pull requests can sometimes take longer due to various factors such as review time, code complexity, and team capacity. The YOLOv5 repository receives a high volume of pull requests, and the team is working diligently to review and merge them as efficiently as possible.

I understand your eagerness to have your pull request merged, and I assure you that we are actively reviewing it. We appreciate your contribution and thank you for your patience. Please rest assured that we will provide an update as soon as possible.

Thank you again for your understanding.

Kind regards,

[deanmark]

@aash1999 hi,

Can you please add the train and val commands, along with the trained weights? I would like to validate the results.
Thanks

[ExtReMLapin]

Tbh, what you call "SODH" is just P2 layer. And it does most of the trick on this paper

[glenn-jocher]

@ExtReMLapin thank you for sharing your insights! It's great to hear your perspective and the importance you attribute to the P2 layer. Your expertise adds valuable context to this discussion. Keep the great feedback coming!

[mahilaMoghadami]

hello,
how to convert this changes (HIC yolo) from yolov5s to yolov5l?
thank you

[glenn-jocher]

@mahilaMoghadami to convert the HIC YOLO changes from YOLOv5S to YOLOv5L, you can adjust the model architecture settings in the YOLOv5 configuration files (yolov5s.yaml and yolov5l.yaml).

In these files, you can modify the "backbone" and "head" sections to match the larger YOLOv5L architecture. Specifically, you would need to update the number of layers, channels, and other architecture-specific parameters to align with YOLOv5L specifications.

After making the necessary adjustments, you can use the YOLOv5L configuration files for training and inference. Be sure to update the command-line arguments for training and evaluation to use the YOLOv5L configuration and model weights.

If you need further assistance, you can refer to the YOLOv5 documentation at https://docs.ultralytics.com/yolov5/ or feel free to ask for specific guidance.

I hope this helps! Let me know if you have any more questions.

[aash1999]

hello, how to convert this changes (HIC yolo) from yolov5s to yolov5l? thank you

you can change the architecture of 5L model by adding CBAM and Involution similar to the one mentioned in this branch. also take care of dimensions and test them. please reach out if you encounter any issues.

Regards

[glenn-jocher]

Hello @aash1999,

Exactly, to adapt the HIC YOLO improvements from the YOLOv5S to the YOLOv5L model, you'll need to integrate the Channel Attention Block (CBAM) and Involution modules into the YOLOv5L architecture. This involves:

1. Editing the YOLOv5L configuration file (typically named yolov5l.yaml) to include the CBAM and Involution layers at the appropriate positions within the network.
2. Ensuring that the dimensions of the layers match the expected input and output shapes, especially when scaling up from YOLOv5S to YOLOv5L.
3. Testing the modified architecture to verify that the network trains correctly and that the performance improvements are consistent with those observed in the smaller model.

If you run into any issues or have further questions, don't hesitate to ask. The community is here to help!

[highquanglity]

@aash1999 I have trained HIC model follow your configure but not get the same result. After 300 epochs, i only got 25 mAP.50. This is model summary:

[glenn-jocher]

Hello @highquanglity,

It looks like your mAP results are lower than expected. A few things to consider:

• Double-check your dataset quality and annotations.
• Ensure that the hyperparameters in hyp.hic-yolov5s.yaml are correctly set for your specific dataset.
• Experiment with different learning rates or training for more epochs.

If the issue persists, could you share more details about your training dataset and the exact command you used for training? This might help in diagnosing the problem more effectively. 🛠️

Keep up the great work, and let's get those numbers up!

[raoufslv]

• Ensure that the hyperparameters in hyp.hic-yolov5s.yaml are correctly set for your specific dataset.

Hello @glenn-jocher,
how exactly, a one can set the hyperparameters correctly to a specific dataset ? (in my case, fire detection)

Thanks.

[glenn-jocher]

Hello @raoufslv,

Great question! To tailor the hyperparameters for your specific dataset, such as fire detection, you can start by training with the default settings to establish a baseline. From there, you can adjust the hyperparameters based on your dataset's characteristics.

For example:

• Learning Rate: Adjust if the model is not converging or is overfitting.
• Batch Size: Use the largest batch size your hardware can handle.
• Augmentation Parameters: Increase augmentation to improve generalization, especially if your dataset is small.

You can also use the Hyperparameter Evolution feature in YOLOv5 to automatically find the best hyperparameters for your dataset. For more details, refer to the Hyperparameter Evolution Tutorial.

Good luck with your fire detection project! 🔥

[Y-T-G]

Thanks @aash1999.

I ported this to YOLOv8 if anyone is interested.
https://gist.github.com/Y-T-G/3b62416a6439a385e743d62f0d0ef842

git clone https://github.com/ultralytics/ultralytics
cd ultralytics
git reset --hard b87ea6ab221ef1e7f7440fe47c6f924a7fce2862
wget https://gist.githubusercontent.com/Y-T-G/3b62416a6439a385e743d62f0d0ef842/raw/1a1defe62c236da11a8302fc40a3ebe3e0410692/YOLOv8-CBAM-Involution.patch
git apply YOLOv8-CBAM-Involution.patch
pip install -e .

from ultralytics import YOLO

# P2 model
model = YOLO("ultralytics/cfg/models/v8/yolov8m-cbam-involution-p2.yaml", task="detect")

# Normal
model = YOLO("ultralytics/cfg/models/v8/yolov8m-cbam-involution.yaml", task="detect")

# Change suffix letter to reload different sizes
model = YOLO("ultralytics/cfg/models/v8/yolov8n-cbam-involution.yaml", task="detect")
model = YOLO("ultralytics/cfg/models/v8/yolov8s-cbam-involution.yaml", task="detect")
model = YOLO("ultralytics/cfg/models/v8/yolov8m-cbam-involution.yaml", task="detect")

[shrutichakraborty]

Hi @glenn-jocher and @aash1999 , I would like to use the CBAM modules you have added to the common.py and the models/hub/yolov5s-cbam-involution.yaml config to train a network for small object detection. However, after I clone the whole yolov5 repo and even pull changes from the master branch, I do not have the CBAM modules in my common.py file and I do not have the models/hub/yolov5s-cbam-involution.yaml file. I wanted to check that these files have indeed been merged with the main branch and are available publicly ? Can you assist me?

[glenn-jocher]

Hi @shrutichakraborty,

The CBAM modules and yolov5s-cbam-involution.yaml configuration are not part of the main YOLOv5 repository. You may need to manually integrate these changes into your local copy. If you need further assistance with this integration, please refer to the documentation or community forums for guidance.

[AmirAliEmami99]

Hi @glenn-jocher and @aash1999, I'm trying to add the CBAM attention module to Yolo architecture at different backbone layers and have several questions.

I added Cbam implementation to common.py and modified the backbone of yolov5s.yaml file (all kernel_size are 3):

# Ultralytics YOLOv5 🚀, AGPL-3.0 license

# Parameters
nc: 80 # number of classes
depth_multiple: 0.33 # model depth multiple
width_multiple: 0.50 # layer channel multiple
anchors:
  - [10, 13, 16, 30, 33, 23] # P3/8
  - [30, 61, 62, 45, 59, 119] # P4/16
  - [116, 90, 156, 198, 373, 326] # P5/32

# YOLOv5 v6.0 backbone
backbone:
  [
    [-1, 1, Conv, [64, 6, 2, 2]],    # 0-P1/2
    [-1, 1, Conv, [128, 3, 2]],      # 1-P2/4
    [-1, 1, CBAM, [128, 128, 3]],    # CBAM after Conv layer
    [-1, 3, C3, [128]],              # 2
    [-1, 1, Conv, [256, 3, 2]],      # 3-P3/8
    [-1, 1, CBAM, [256, 256, 3]],    # CBAM after Conv block for mid-level features
    [-1, 6, C3, [256]],              # 4
    [-1, 1, Conv, [512, 3, 2]],      # 5-P4/16
    [-1, 1, CBAM, [512, 512, 3]],    # CBAM after Conv block for high-level features
    [-1, 9, C3, [512]],              # 6
    [-1, 1, Conv, [1024, 3, 2]],     # 7-P5/32
    [-1, 1, CBAM, [1024, 1024, 3]],  # CBAM for deep features
    [-1, 3, C3, [1024]],             # 8
    [-1, 1, SPPF, [1024, 5]],        # 9
  ]

and this is the modification of the parser in yolo.py:

if m in {
            Conv,
            GhostConv,
            Bottleneck,
            GhostBottleneck,
            SPP,
            SPPF,
            DWConv,
            MixConv2d,
            Focus,
            CrossConv,
            BottleneckCSP,
            C3,
            C3TR,
            C3SPP,
            C3Ghost,
            nn.ConvTranspose2d,
            DWConvTranspose2d,
            C3x,
            CBAM
        }:
            c1, c2 = ch[f], args[0]
            if c2 != no:  # if not output
                c2 = make_divisible(c2 * gw, ch_mul)

            print(f"layer args before: {args}")  # Debug to print arguments
            args = [c1, c2, *args[1:]]
            print(f"layer args after: {args}")  # Debug to print arguments

But i face the bellow error:

                 from  n    params  module                                  arguments                     
layer args before: [64, 6, 2, 2]
layer args after: [3, 32, 6, 2, 2]
  0                -1  1      3520  models.common.Conv                      [3, 32, 6, 2, 2]              
layer args before: [128, 3, 2]
layer args after: [32, 64, 3, 2]
  1                -1  1     18560  models.common.Conv                      [32, 64, 3, 2]                
layer args before: [128, 128, 3]
layer args after: [64, 64, 128, 3]
Initializing SpatialAttention with kernel_size=128
Traceback (most recent call last):
  File "/content/drive/MyDrive/thesis/yolov5/train.py", line 986, in <module>
    main(opt)
  File "/content/drive/MyDrive/thesis/yolov5/train.py", line 688, in main
    train(opt.hyp, opt, device, callbacks)
  File "/content/drive/MyDrive/thesis/yolov5/train.py", line 216, in train
    model = Model(cfg or ckpt["model"].yaml, ch=3, nc=nc, anchors=hyp.get("anchors")).to(device)  # create
  File "/content/drive/MyDrive/thesis/yolov5/models/yolo.py", line 245, in __init__
    self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch])  # model, savelist
  File "/content/drive/MyDrive/thesis/yolov5/models/yolo.py", line 476, in parse_model
    m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args)  # module
  File "/content/drive/MyDrive/thesis/yolov5/models/common.py", line 170, in __init__
    self.spatial_attention = SpatialAttention(kernel_size)
  File "/content/drive/MyDrive/thesis/yolov5/models/common.py", line 127, in __init__
    assert kernel_size in (3, 7), 'kernel size must be 3 or 7'
AssertionError: kernel size must be 3 or 7

1. I defined the CBAM module with [128, 128, 3] args but the parser output is [64, 64, 128, 3] which set the kernel size to 128. why?

2. i can't find the relation between args in yaml file and the args after parse function, for example:

first two layers args in yaml file:

    [-1, 1, Conv, [64, 6, 2, 2]],    # 0-P1/2
    [-1, 1, Conv, [128, 3, 2]],      # 1-P2/4

first two layers args in runtime:

                 from  n    params  module                                  arguments                     
  0                -1  1      3520  models.common.Conv                      [3, 32, 6, 2, 2]            
  1                -1  1     18560  models.common.Conv                      [32, 64, 3, 2]

3. How do we manage each layer's arguments to avoid errors like this?