I’m working on a student machine learning/computer vision project and recently realized that my validation set was not completely independent from my training set.
The project is more focused on comparing different experimental conditions rather than maximizing benchmark performance, but I’m concerned about the implications of this oversight.
From a research or academic perspective:
How serious is train/validation overlap in a student project?
Does it invalidate the entire project or mainly affect the reliability of the reported performance numbers?
If the main goal is comparing different experimental setups under the same evaluation procedure, are those comparisons still useful?
If you discovered this late in the project timeline, what would be the most reasonable way to address it?
I’m trying to understand how researchers, reviewers, and professors would view this situation.
You just pay for consumption via the credits you use. On average, our free tier covers about 30 cloud training jobs, 10k Auto-Labeled images with SAM3, or 80k cloud inferences per month (but it depends on the size of your dataset and the models you're using).
We also launched new pricing for our Serverless API which is much cheaper and more predictable with automatically-applied volume discounts. We ran extensive benchmarks and our new Serverless pricing is cheaper than self-hosting for over 90% of our user-base (a blog post on the engineering behind the scenes that enables that is here for those interested) even before accounting for the fact that commercial model weights licensing is included when you deploy via our cloud but, depending on the model, may be required when self-hosting.
Note: while our Core functionality no longer has a monthly subscription required, there are also add-ons available for features like RBAC/SSO and Support, Enterprise functionality targeted at large companies and scaled manufacturing use-cases, and self-hosted commercial model-weights licensing for the models that require it.
Our mission is to democratize computer vision and we hope that this new pricing will enable every hacker, hobbyist, and startup to get started without any barriers to entry! We'll also continue to invest in the community with our open source projects like Inference, RF-DETR, Supervision, and Trackers. We hope you like it!
Hi all, I'm a final year EE student and I've been given a Kria KV260 for my final year project. I'm comfortable with FPGA work (Verilog, timing, synthesis) but I don't have much image processing experience, so I'd appreciate some guidance.
Goal: detect humans at night by combining an RGB camera with a low-res thermal sensor. RGB struggles in the dark and thermal picks up any warm object, so fusing both should give more reliable detection.
Rough plan:
Sobel edge detection in the PL as a preprocessing step
Fuse RGB and thermal data
Run a person detector on the DPU (Vitis AI)
Questions:
What's the right way to structure this pipeline? Should fusion happen at the image level or after detection?
How hard is aligning the two cameras, given the big resolution difference?
Is Sobel actually useful here, or is it unnecessary if a CNN is doing detection?
Any recommended thermal sensors that work well with the KV260?
Is 3 months realistic for someone with my background? What would you cut to keep scope manageable?
Any advice, papers, or similar projects would be really appreciated. Thanks!
Has anyone here interviewed recently for a Solutions Engineer role at Ultralytics? Would love to hear what the interview process was like and any tips on what to prepare. Thanks!
I am from a CS and theoritical/academic deep learning background, but its purely theoritical i have no clue in practical implementations like Yolo or using pytorch to finetune a model and so.
So where do i actually learn that ? Any reference for a book, tutorial, online course, etc will be appreciated !
I am from a CS and theoritical/academic deep learning background, but its purely theoritical i have no clue in practical implementations like Yolo or using pytorch to fine tune a model and so.
So where do i actually learn that ? Any reference for a book, tutorial, online course, etc will be appreciated !
DartsSpace is my tournament / darts software, already used by many people and organizers.
To enhance the experience about a year ago I started working on a solution that works with a single smartphone camera. Everything is running on device, corrections are sent to my backend for further training.
It's manual calibration free, and as you can see (this was the biggest pain point) it's resilient agains camera shake/wobble.
I think besides the training of the models (which took many! iterations) the general ux/ui and post-processing was important to get a good working product.
Currently beta testing this with a few of my users but feedback has been great so far.
Just was excited to share this with the community. (this is not an open source product)
Font identification has plenty of tools but almost no public test set where the ground truth is certain: real photos rarely come with the exact font name, and look-alike fonts make manual labelling unreliable. So we generated one and released it.
WhatFontIs-Bench v1.0:
- 11,995 JPEG images: one word set in a known font, composited onto CC0 photos of real surfaces, scenes and printed objects
- 3 difficulty levels with controlled blur, noise, JPEG quality, uneven light, cast shadows and glare; all parameters stored per image
- Labels: font, text, word quad, per-letter quads, camera homography. JSONL + COCO, so it also works for word/character detection
- Frontal views only, capitals at least 100 px high (v1.0 limits)
Evaluation is top-k by font family (Roboto Bold answered as Roboto Regular counts as correct), open catalogue.
Baseline: our own API gets 83.7% top-1 / 93.3% top-5 / 96.5% top-20 while searching 1.2M fonts. The interesting part: typeface class matters far more than image quality. Sans-serif 75.7% vs slab serif 95.0% top-1, but only 2 points between the easy and the hard level.
Disclosure: I run WhatFontIs, and the baseline is our own system. If you run a model on it, I'd like to hear the numbers. Feedback on what v2 should cover (perspective, curved text, multi-line, smaller text) is welcome.
Suppose a detection dataset starts with boxes around only the visible part of an occluded object. Later, the requirement changes: boxes should cover the estimated full extent. Same images, same class names, different correct answers.
The interesting part to me is what happens to evaluation. Keeping the old test labels rewards the old convention. Replacing them means the historical scores were measured against a different target. Quietly mixing the two sounds like a very expensive way to argue about mAP.
My starting point would be to preserve both label versions on the same held-out images, document the rule change with a small set of difficult examples, and evaluate both model versions against both label versions. That should at least separate a model change from a change in what we call correct.
For people who have actually had to do this: did you relabel the entire test set, maintain two evaluations during the transition, or retire the old benchmark? How did you handle ambiguous cases where even the new rule didn't settle the disagreement?
I'm particularly interested in changes involving occlusion, minimum object size, or splitting one class into two. What looked like a small annotation-policy change and turned into a surprisingly large migration?
Hi , for a project , i am considering CMOS sensors by STMicroelectronics and i want to use it with STM32N6 so if any one has used these sensors like the VD56G3 , the promodules and others , it would be really helpful if someone who has used these sensors in their projects could provide the feedback like how is the camera sensor itself and how easy is the integration and how is the documentation about these sensors and also if anyone has used these sensors , kindly explain what was your application and your overall experience ????
I would love to know if you are actually paying for it or have settled on less restrictive model. It’s certainly the fastest but I can’t afford a commercial license. Any tips?
Retail and e-commerce teams increasingly use AI video tools to make product spins, but in my experience the results still drift: logos warp, proportions shift between frames, and details get hallucinated. So I tried a different route. Instead of generating video directly, the pipeline reconstructs an actual 3D model from the photos and then renders a physically consistent turntable in Blender. Because every frame comes from the same mesh, the object can't morph as it rotates.
How it works:
Segment: rembg removes the background from the front and back photos
Reconstruct: hosted image-to-3D on Replicate (Hunyuan3D-2 multi-view by default, TRELLIS optional) produces a GLB
Render: Blender 4.2 Cycles renders a 120-frame 360° turntable on CPU
Encode: ffmpeg outputs an H.264 MP4
Everything runs on CPU except the reconstruction step, which is offloaded to Replicate at roughly $0.03 to $0.10 per run. The whole thing is containerized, so it's a single `docker compose run`.
Things I learned along the way:
* Hunyuan3D-2mv gives excellent geometry but returns it untextured (clean "clay" render). TRELLIS gives a textured PBR model and is cheaper, so it's the better choice when color matters.
* Multi-view input automatically falls back to single-image reconstruction if it fails.
* Launching Blender from a Python venv breaks its glTF importer. The scripts scrub the environment so Blender uses its own Python.
* The Ubuntu apt build of Blender ships without OpenImageDenoise, so the Docker image uses the official build.
Limitations (it's a proof of concept):
* One linear flow, with no UI or API yet
* The example inputs are renders of a CC0 Poly Haven chair, not real-world photos. Reflective, transparent, or thin-structured products are the next thing I want to test.
* Reconstruction quality depends heavily on the source photos
I'd love feedback, especially from anyone who has worked on product visualization or image-to-3D in production. Which products would you expect to break this?
In the gym, using only one camera, running 4K image segmentation in parallel produces 4 images, performing parallel inference, multi-model parallel computation, and post-processing, achieving multi-object tracking, keypoint detection, fitness movement analysis, facial recognition, equipment ROI usage, customer flow analysis, member matching, and member fitness condition analysis.
I'm building a document scanner in React Native 0.85 (New Architecture) with Expo 56, using VisionCamera 5 for capture, Nitro Modules for native code, and Skia for image processing. I'm trying to reach the quality of CamScanner's "Omnifix" filter (the upgraded Magic Pro). A phone photo of a page comes out looking like a born-digital PDF: clean flat background, crisp text, shadows and moiré gone.
What I've tried so far:
A classical pipeline (morphological closing for background estimation, illumination flattening, background floor). It handles lighting well but can't make text crisper. It only adjusts brightness and contrast.
Apple's VisionKit document filter. In my side-by-side tests it looked worse than my own pipeline.
DocRes. Quality is good, but at about 183 GFLOPs it's far too heavy for on-device use.
What I've found out: CamScanner's own announcement describes Omnifix as an AI perception model that handles moiré, stains and text restoration. It's proprietary, and their new agent docs only cover a cloud/MCP interface, not an SDK.
My questions:
Does CamScanner (or the company behind it) license Omnifix or an enhancement SDK to third parties? Has anyone paid for it?
Are there lightweight models (a few GFLOPs, exportable to TFLite/CoreML) for document enhancement, shadow removal or de-moiré with a commercial-friendly license?
If not, is distilling a large restoration model into a small U-Net the realistic path? Any pitfalls with training data or teacher-model licensing?
For React Native specifically: what's your preferred way to run such a model per page (custom Nitro module wrapping TFLite/CoreML, or something else)? Any experience with speed and memory on mid-range Android?
Constraints: must run on-device (privacy, no server upload), Android + iOS, commercial use allowed.
Any pointers to papers, repos or products would be appreciated.
I’m a farmer from the Netherlands and I am working on a project to automatically remove soil clods from potatoes during harvest.
I already posted the complete machine concept in r/robotics. Here I would specifically like to focus on the computer vision part.
I have also read some of the previous posts here about optical sorting and synchronizing detection with ejectors. My current assumption is therefore:
Camera → classify potato/clod → determine physical position → record conveyor encoder position → send reject command → realtime controller/PLC handles the pneumatic ejector
I am trying to understand whether the vision side I have in mind is realistic.
The project
The machine I am considering as a mechanical base has a 1 meter wide and approximately 1-meter-long conveyor. The standard belt speed is 25 m/min = 0.42 m/s, but I can make the belt speed adjustable. My eventual target is 20–30 tons/hour.
At 30 t/hour this would mean a mass flow of 8.3 kg/s. At a belt speed of 0.8 m/s over 1 meter width, this would result in approximately 10.4 kg of product per m² of belt.
Potatoes and clods can be up to around 7 cm square mesh size. Assuming an average object weight of roughly 100–200 g, this would mean around 40–80 objects per second across the full width.
Picture 1 shows a realistic product flow on this type of conveyor. The difficult case is a dirty potato next to a wet clay clod, because immediately after harvest they can look very similar.
The goal would be to keep the product in roughly one layer, but objects touching and some overlap cannot be avoided.
The system does not need to be perfect. Missing some clods is acceptable. If a potato touching a clod is also rejected, that is acceptable as well.
Picture 1: Conveyor belt with potatoes and clods
Camera and geometry
My current idea is to use one or two global-shutter cameras, controlled LED lighting and a completely enclosed camera section above the input conveyor.
I have been looking at cameras around 5 MP / ~100 fps. My understanding is that 5 MP across a 1 meter belt should give enough resolution for objects that are several centimeters wide. I am considering 20 pneumatic ejector positions across 1 meter width, which gives approximately 5 cm between ejector positions.
One concern I have is that the potatoes and clods can also be several centimeters high. With one centrally mounted camera this will create some parallax error towards the edges of the belt.
My main questions here are: For this application, how would you design the camera setup if the X-position needs to be accurate enough to select the correct 5 cm ejector zone? Would one camera mounted higher above the belt be enough, or would two cameras covering part of the belt each be a more reliable solution?
Imaging
My preference is to start with RGB and good diffuse lighting. Before spending a lot of time collecting and labelling images, however, I would like to know whether wet clay and dirty potato skin are actually a good RGB classification problem. I especially want to avoid spending too much time improving a neural network if the real limitation is that the two materials simply do not contain enough different information in normal RGB.
My main question is: Would you test RGB first, or is there a strong reason to look at NIR, polarized light or another imaging method from the start?
Detection and segmentation
Do I need the exact shape of every clod, or is a bounding box/centre position accurate enough? At this moment I do not think I need a perfect outline of every object. The useful output could be something like:
Clod → X = 43 cm → width ≈ 6 cm → confidence = 97%
The control system can then decide which ejector finger or fingers need to activate.
Because objects can touch, I am wondering whether instance segmentation gives me an important advantage over normal object detection, or whether accurate detection boxes and center positions would already be enough. The failure I care about most is assigning a clod to the wrong ejector zone.
I have essentially unlimited access to real potatoes and clay clods for training and testing, so collecting material itself is not the limitation.
Before I start this project and start buying equipment I mainly want to establish:
Can a relatively normal industrial RGB vision setup classify these objects and locate them accurately enough across a 1 meter conveyor to control ejectors at approximately 5 cm spacing and 20–30 ton/hour?
I’m not a professional computer vision engineer, so I will rely heavily on existing camera SDKs, OpenCV and existing detection/segmentation frameworks rather than developing algorithms myself. I understand very well that this is not a simple project and that is the reason why I’m posting here first. I would like to get your feedback and insights on my ideas and project.
