Poor results in the detection of long and thin objects

[Ichiruchan]

Hello everyone, I am currently facing a challenge. My goal is to segment all elongated objects in the image, which looks like this:

The red lines represent the objects. The image size is 1400 pixels in width and 300 pixels in height. The bounding box for segmentation has a width ranging from 3 to 15 pixels and a height ranging from 20 to 40 pixels. I have noticed that the detection results for objects with a width smaller than 8 pixels are poor when using the yolov8x-seg.pt model.

1.I added the p2 section to the YAML file, but the results remained unsatisfactory. When I forced the bounding box width to be greater than 8 pixels, the results improved significantly. However, this approach requires extensive code changes in YOLO.
2. I attempted to split the image into multiple sections, but result is still bad.
3. I modified the YAML file to double the size of the feature maps, but it resulted in running out of GPU memory. I also believe that the channels in the shallower layers are insufficient:

# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLOv8-multitask multitask model. For Usage examples see https://docs.ultralytics.com/tasks/multitask

# Parameters
nc: 80 # number of classes
kpt_shape: [17, 3] # number of keypoints, number of dims (3 for x,y,visible)

scales: # model compound scaling constants, i.e. 'model=yolov8n-mltitask.yaml' will call yolov8-seg.yaml with scale 'n'
  # [depth, width, max_channels]
  n: [0.33, 0.25, 1024]
  s: [0.33, 0.50, 1024]
  m: [0.67, 0.75, 768]
  l: [1.00, 1.00, 512]
  x: [1.00, 1.25, 512]

# YOLOv8.0n backbone
backbone:
  # [from, repeats, module, args]
  - [-1, 1, Conv, [32, 3, 1]] # 0
  - [-1, 1, Conv, [64, 3, 2]] # 1
  - [-1, 3, C2f, [64, True]]  # 2
  - [-1, 1, Conv, [128, 3, 2]] # 3
  - [-1, 3, C2f, [128, True]] # 4
  - [-1, 1, Conv, [256, 3, 2]] # 5
  - [-1, 6, C2f, [256, True]]  # 6
  - [-1, 1, Conv, [512, 3, 2]] # 7
  - [-1, 6, C2f, [512, True]] # 8
  - [-1, 1, Conv, [1024, 3, 2]] # 9
  - [-1, 3, C2f, [1024, True]] # 10
  - [-1, 1, SPPF, [1024, 5]] # 11

# YOLOv8.0n head
head:
  - [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 12
  - [[-1, 8], 1, Concat, [1]] # 13
  - [-1, 3, C2f, [512]] # 14

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]] #15
  - [[-1, 6], 1, Concat, [1]] # 16
  - [-1, 3, C2f, [256]] # 17

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 18
  - [[-1, 4], 1, Concat, [1]] # 19
  - [-1, 3, C2f, [128]] # 20

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]] # 21
  - [[-1, 2], 1, Concat, [1]] # 22
  - [-1, 3, C2f, [64]] # 23

  - [-1, 1, Conv, [64, 3, 2]] # 24
  - [[-1, 20], 1, Concat, [1]] # 25
  - [-1, 3, C2f, [128]] # 26

  - [-1, 1, Conv, [128, 3, 2]] # 27
  - [[-1, 17], 1, Concat, [1]] # 28
  - [-1, 3, C2f, [256]] # 29

  - [-1, 1, Conv, [256, 3, 2]] # 30
  - [[-1, 14], 1, Concat, [1]] # 31
  - [-1, 3, C2f, [512]] # 32

  - [-1, 1, Conv, [512, 3, 2]] # 33
  - [[-1, 11], 1, Concat, [1]] # 34
  - [-1, 3, C2f, [1024]] # 35

  - [[23, 26, 29, 32, 35], 1, Segment, [nc, 32, 256]]

4. I doubled the channels of backbone layers. The result didn't become well.
5. I doubled the image width while keeping the height the same. The result become better, but not enough.

Could you give me some suggestion to improve the result.

[UltralyticsAssistant]

👋 Hello @Ichiruchan, thank you for sharing your detailed experience and reaching out to the Ultralytics community 🚀! We understand how tackling specific cases like segmenting elongated and thin objects can be uniquely challenging.

We suggest reviewing our official Docs for insights and tips—many users have found our Model Training Tips and examples helpful for improving results. There might also be strategies applicable to your use case within our Guides section.

Suggestions to Move Forward

If this is a 🐛 Bug Report or something related to model behavior, please create a minimum reproducible example showcasing the issue you're experiencing. Providing example images or logs would be especially helpful in understanding the problem.

In addition:

• Let us know more specifics about your dataset, such as image variability, object distribution, and augmentation techniques, if applicable.
• Verify you are using the latest version of ultralytics. Upgrade as follows to ensure you have the newest updates and bug fixes:

  pip install -U ultralytics

Community Resources

For further help, feel free to share your findings or brainstorm ideas with our vibrant Ultralytics community:

• Join us on Discord for real-time chats 🎧.
• Engage in deeper discussions on our Discourse forum.
• Explore our Subreddit for additional advice from enthusiasts worldwide.

Verified Environments

To rule out any potential setup issues, consider running your experiments in one of our verified environments where dependencies like CUDA, Python, and PyTorch come pre-installed:

• Notebooks: Use free GPUs here:
• Cloud Platforms: Leverage Google Cloud or AWS.
• Docker: Check out our Docker Quickstart Guide .

Status Verification

Check the current status of our CI tests here:

Green signals indicate all functionality, including segmentation tasks, is verified across supported platforms and modes.

This is an automated response to guide you in the right direction 🌟. An Ultralytics engineer will also address your query soon to provide further assistance.

[Y-T-G]

Objects that are very long and thin get removed.

ultralytics/ultralytics/data/augment.py

Line 1262 in d170f1a

def box_candidates(self, box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16):

You can try modifying it to not remove those. And tey using mask_ratio=1 during training.

[glenn-jocher]

@Ichiruchan you're welcome! Let us know how it goes with the suggested modifications, and feel free to share any updates or further questions. The YOLO community and Ultralytics team are here to help!

[Ichiruchan]

I tried it, but it doesn't work. I printed the return value, and there are no false. I believe the model may not be suitable for detecting features with very small widths. The feature map size is 1/4, 1/8, 1/16 and 1/32.

[glenn-jocher]

Thank you for sharing your findings. For detecting very small-width features, consider increasing the image resolution during training and inference (imgsz) or experimenting with a custom model architecture that incorporates finer feature maps (e.g., 1/2 or 1/4 scales). Additionally, tweaking the conf and iou thresholds might help refine detections. Let us know how it goes!

[Ichiruchan]

I changed the image size to four times the original dimensions, but it didn't work. However, when I only doubled the width, it worked well. I believe that my width-to-height ratio is less than 100.

[glenn-jocher]

@Ichiruchan thank you for sharing your observations. Doubling the width likely improved results because it provided greater resolution for the long, thin features relative to their orientation. If this approach works well for your case, it seems like a practical solution. Let us know if further adjustments or insights are needed!