←
Satya Mallick4,062 posts
Satya Mallick avatar

Satya Mallick

@LearnOpenCV

CEO, OpenCV.org. CEO, BigVision.AI Entrepreneur. Ph.D. ( Computer Vision & Machine Learning ). Author: LearnOpenCV.com

◷ Joined June 2008▧ 1,444 Photos and videos
Satya Mallick avatar
Satya Mallick@LearnOpenCV·7h

The surprising part was not 99%. It was the AI scoring 0% without the camera’s scale. Can AI See Glass Better Than a Depth Camera? Only with the camera’s help. We tested Depth Anything V2 (large, metric indoor) on 113 real ClearGrasp scenes captured with Intel RealSense depth cameras. A glass pixel counted as correct if it was within 5% of the reference depth; missing readings counted as wrong. On one clear glass dish: camera 7%, AI + camera scale 99%. Across all 113 scenes: camera 32%, AI + camera scale 78%. The scaled AI did better in 101 scenes. Without scaling, the AI scored 0% on the glass in this experiment. We set its scale using the camera’s non-glass depth, not the reference depth. Away from glass, the camera was better: 98% within the tolerance, versus 83% for the AI. The camera supplies reliable scale; the AI helps recover shape. Use them together. Scope: one model, 113 lab scenes from 2020. Reference depth came from the same camera after glass objects were replaced with opaque copies. This is an experiment, not a benchmark or a claim about every camera or model. Data: ClearGrasp real test scenes, Sajjan et al., “ClearGrasp: 3D Shape Estimation of Transparent Objects for Manipulation”, ICRA 2020 (Google and collaborators). sites.google.com/view/clearg… Model: Depth Anything V2, Yang et al., 2024, Metric Indoor Large. Intel RealSense is a trademark of Intel. No endorsement implied. #ComputerVision #DepthEstimation #Robotics

2:54
Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 17

A depth map can get the room’s shape right and its distances wrong. Calibrate at the zoom you actually use. YOLO26 Depth Estimation: Can You Trust the Meters? We tested YOLO26n-depth on 654 indoor NYU Depth V2 photos to see how much we could trust its distances. A simulated 2x zoom made the same surfaces read about 46% closer. One wall, 2.4 m away, read 2.79 m on the full photo and 1.37 m after zooming. Rescaling each prediction to the ground truth before grading gave 78% within 25% of the true distance. Grading the raw meters gave 53%. Out of the box, readings were about 18% too far, with roughly 67 cm average error. One scale correction fitted from 150 ground-truth points on 30 training photos reduced the bias to 2%. But applying that correction to zoomed photos made them worse: 45% too close. Fitting on zoomed training photos instead brought the bias to 6%. Takeaway: check the distances on your own camera and calibrate at the zoom you will actually use. Scope: one nano model, one indoor dataset, zoom simulated by centre-cropping and resizing. NYU Depth V2 may be part of the training mix, so treat this as an experiment, not a benchmark. Test photos were not used to fit the calibration. Data: NYU Depth V2, Silberman, Hoiem, Kohli and Fergus, “Indoor Segmentation and Support Inference from RGBD Images”, ECCV 2012. Model and calibration guidance: Ultralytics YOLO26 depth. No endorsement implied. #ComputerVision #DepthEstimation #YOLO26

5:04
Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 17

73 defects flagged. 27 missed. And a flagged image is not the same as a well-localised defect. We built a PatchCore-based detector from 492 photos of good capsules in VisA, and set its cut-off with 50 separate good photos. On 160 test images, it flagged 73 of 100 defective images, missed 27, and raised zero false alarms on 60 good images. Finding a defective image is only part of the job. On the opening capsule, the marked area covered about 26% of the labelled defect. Another example peaked on the defect but stayed below the cut-off. Check what the detector misses, what it flags by mistake, and whether it points at the defect. Development experiment, not a production test: we reused the test set while building the method. These numbers do not predict factory performance. The two examples were selected after results to show a partial hit and a miss. Method: PatchCore (Roth et al., 2022), ImageNet-pretrained WideResNet-50, 2% coreset. Images/annotations: VisA, Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang and Onkar Dabeer, SPot-the-Difference (2022). Resized and annotated with our experiment outputs. CC BY 4.0. No endorsement implied. github.com/amazon-science/sp… creativecommons.org/licenses… #AnomalyDetection #ComputerVision #OpenCV

2:17
Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 16

A few years back I used to pay $5/mo for @evernote - that's $60/year. Over time I started paying $120/year which is understandable. This year they upgraded me to $249/year automatically! Yes, they sent me emails well in advance but who reads these emails. My account would have been fine without the upgrade, and I should have been asked to proactively upgrade not an automatic upgrade. Anyway, I asked my agent to create a backup of my notes, move them to Google Drive, create a human readable version of it, and ask for a refund. They $249 refund is on its way. Ironically, Evernote's automatic upgrade was for AI features I never used, and my refund was made possible by AI that I actually use.

Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 16

Finding every object does not guarantee the right count. Our shelf detector found 30 cans plus a box around two of them. IoU let it survive NMS; containment flagged it. One photo/model, not a benchmark. Photo: AustinNeil, Wikimedia Commons, CC BY 4.0. Cropped and annotated. No endorsement. commons.wikimedia.org/wiki/F… creativecommons.org/licenses…

1:04
Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 13

One second of Kohli's flick shot. Six AI models. One 3D ribbon. SAM grabbed the stumps too. The fix: ask for wicket and subtract. Relative depth, hand-scaled: a sketch, not a measurement. One clip, not a benchmark. Footage: Cricket Australia / Fox Cricket, for analysis and commentary. No endorsement. Notebook: learnopencv.com/decoding-vir… #ComputerVision #ViratKohli

7:25
Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 13

Inverse Cramer predicts Jensen will be back!

How venal, how disgusting are some that they are willing to look away from a ceo who is having a relationship with a subordinate simply because he is brilliant or a money-maker. No one was a bigger defender of Jensen than I was but there are rules about sex with subordinates. Just because someone is brilliant doesn't mean he can do what he wants

Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 13

A dark background can hide a bad hair cutout. Our PyMatting demo kept finer wisps with fractional alpha; rounding to 0/1 made blobs. Check dark hair on a light background. One portrait with manual guidance, not a benchmark. Photo: Tom Morbey/Unsplash. No endorsement.

0:54
Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 12

Would you trust the sharper image? We shrank a NASA photo and rebuilt it with AI and bicubic resizing. AI looked better; bicubic matched the original more closely. One photo/model, not a benchmark. Photo: NASA/Robert Markowitz, via Wikimedia Commons. No endorsement.

1:28
Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 11

81 milliseconds apart on the arrow release. 714 milliseconds apart on the start of the draw. The boundary definition matters. We gave Gemini 4 Carbon a 14-second archery clip and checked its timestamps against our own frame-by-frame marks. Release: Carbon said 11.000 s; we marked 10.919 s. That is 81 ms apart, about 2 frames at 25 fps. Start of the draw: Carbon said 6.000 s; we marked 5.286 s. That is 0.714 s apart, about 18 frames. On a second clip, our own two marking passes were 7 ms apart on the release and 83 ms apart on the start of the draw. Sharp events are easier to pin down than gradual ones. Before labelling video, define exactly where each phase begins. Scope: one model, two clips, manual markers. Our marks are visual estimates, not frame-perfect ground truth. This is a small experiment, not a benchmark. Footage: Mikhail Nilov via Pexels (Pexels License): pexels.com/video/side-view-o… “Field Archery” by Simon Allen, Wikimedia Commons, CC BY 3.0: commons.wikimedia.org/wiki/F… creativecommons.org/licenses… Changes: excerpted, slowed, frames extracted, audio removed, annotations and commentary added. No endorsement by the creators or people shown is implied. Model: Gemini 4 Carbon (Google), used through the Gemini app. #ComputerVision #VideoUnderstanding #Gemini

1:06
Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 10

A library upgrade can change your detections without changing your model. We compared Ultralytics 8.4.207 and 8.4.208 using identical YOLO11n weights and seven images. Four images changed their box counts. The cause: prediction's default NMS IoU moved from 0.7 to 0.5. One explicit argument—iou=0.7—reproduced the old boxes, scores and classes exactly on all seven images. Fewer boxes did not mean we found missing people: the six removed boxes were four duplicates and two broad boxes spanning overlapping players. This was a small controlled experiment, not an accuracy benchmark. The practical lesson: own your thresholds, pin dependencies, and keep regression images. Footage: SoccerTrack v2 by Atom Scott et al., CC BY 4.0, cropped/upscaled with detection overlays. atomscott.github.io/SoccerTr… creativecommons.org/licenses… Tie close-up: cropped Ultralytics zidane.jpg sample with overlays. Release notes: github.com/ultralytics/ultra… YOLO11n by Ultralytics. No endorsement implied. #ComputerVision #MLOps #YOLO

6:00
Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 10

We gave Mirage, a 22 MB offline vision model, 10 real signs. Plain words: 6/6 right. Accented words: 0/8. ARRÊT became ARRET on four stop signs; Straße became Strabe. Hindi and Telugu on two Indian railway boards: 0/4. Test in your users' language before you ship. Small experiment, not a benchmark: one model, three runs per image. Two boards cannot establish performance across Indian languages. Images: Wikimedia Commons, CC BY-SA, by Mike Fitzsimon, Tony Webster, B3251, Malimage, BrokenSphere, Ana Paula Hirama, Ben AlHürxe and Adityamadhav83. Demo image: Cactus Compute / COCO. No endorsement implied. #ComputerVision #OCR #EdgeAI

1:32
Khoa Vu avatar
⟳ Satya Mallick retweeted
Khoa Vu@KhoaVuUmn·Nov 9

The fact that two Huangs are involved in an affair, but only one takes the blame is commonly known as Jensen's Inequality.

Satya Mallick avatar
Satya Mallick@LearnOpenCV·Nov 9

AI & Computer Vision News (Nov 2–6): Demo vs Real Data Seven stories, from Sekai's event memory to annotator-free object detection. Authors report 60.8 AP on COCO and 47.4 on LVIS, zero-shot, without human labels on target data. We haven't reproduced these results; the foundation models used human supervision. Sources and footage: Sekai (arXiv 2611.06685); huggingface/lerobot and allenai/toyblocks (Apache-2.0); hanyucc/trainspotting (2611.06849); OpenBMB/MiniCPM-OCR (Apache-2.0); VLALagBench (2611.03103); Annotator-Free Object Detection (2611.06195); OpenAI and Google. Office and factory scenes are illustrations made with Gemini Omni 1.2 Flash. No endorsement implied. #ComputerVision #AINews

5:59