1 00:00:03,510 --> 00:00:09,210 You're the World is a zero shot object detection model, which means we can do object detection without 2 00:00:09,210 --> 00:00:11,160 training the object detection model. 3 00:00:11,160 --> 00:00:17,550 So in YOLO world, all we have to do is prompt the model by specifying the list of classes that we are 4 00:00:17,550 --> 00:00:20,580 looking for and in an image or video. 5 00:00:20,580 --> 00:00:21,720 And that's it. 6 00:00:21,720 --> 00:00:25,170 No training is required in case of YOLO world. 7 00:00:25,260 --> 00:00:31,740 So YOLO world is this is designed to solve a limitation of existing zero shot object detection model 8 00:00:31,740 --> 00:00:33,090 which is speed. 9 00:00:33,090 --> 00:00:34,890 Or you can say latency. 10 00:00:34,890 --> 00:00:36,000 So what is latency? 11 00:00:36,030 --> 00:00:42,390 Latency is basically the time taken by the object detection model to do object detection when an input 12 00:00:42,390 --> 00:00:42,900 image. 13 00:00:42,900 --> 00:00:49,320 So in case of YOLO world we solve this limitation of existing zero shot object detection models. 14 00:00:49,590 --> 00:00:57,180 So the other state of the art object zero shot object detection models like grounding dino uh, use 15 00:00:57,180 --> 00:01:00,870 transformer based architecture, which is typically slow architecture. 16 00:01:00,870 --> 00:01:07,290 But in case of YOLO world, YOLO world is being designed using faster CNN based YOLO architecture. 17 00:01:07,290 --> 00:01:11,790 So YOLO world is being designed using faster CNN based YOLO architecture. 18 00:01:12,270 --> 00:01:16,290 So the downside of Granadino is its speed. 19 00:01:16,290 --> 00:01:21,090 So the downside of grounding 0 or 0 shot object detection model is its speed. 20 00:01:21,090 --> 00:01:27,480 So the time like grounding model takes around one second to process a single image, which is pretty 21 00:01:27,480 --> 00:01:27,840 slow. 22 00:01:27,840 --> 00:01:32,760 If you are thinking about processing live video streams using Grounding Dino. 23 00:01:36,750 --> 00:01:40,350 The following chart is being taken from the Yellow World paper. 24 00:01:40,830 --> 00:01:47,550 So from the chart we can see that yellow world maintains almost the same accuracy, and it is 20 times 25 00:01:47,550 --> 00:01:52,890 faster and five times smaller than the leading zero shot object detection models. 26 00:01:53,280 --> 00:02:01,140 According to the YOLO world paper, the small version of the YOLO YOLO world model achieves up to 74.1 27 00:02:01,140 --> 00:02:06,210 fps on Nvidia V 100 GPU, which is quite impressive. 28 00:02:09,010 --> 00:02:16,420 Traditional object detection models such as faster R-cnn, SSD, and YOLO models are designed to detect 29 00:02:16,420 --> 00:02:19,810 objects within a predefined set of categories. 30 00:02:20,170 --> 00:02:26,650 For example, models which are trained on Coco dataset are limited to limited to 80 categories. 31 00:02:26,650 --> 00:02:32,620 So if you want a model to detect new objects that does not exist in the Coco data set. 32 00:02:32,620 --> 00:02:37,480 For example, if I want to detect a pistol or if I want to detect personal protective equipment, I 33 00:02:37,480 --> 00:02:40,240 if I want to do personal protective equipment detection. 34 00:02:40,240 --> 00:02:46,210 So then what we need to do, we need to, uh, create a data of personal protective equipment detection 35 00:02:46,390 --> 00:02:50,230 where we have the images of, like, different personal protective equipment. 36 00:02:50,230 --> 00:02:55,540 And if you want to do, uh, pistol detection, I need to create a data set where we have different 37 00:02:55,540 --> 00:02:56,770 images of pistols. 38 00:02:56,770 --> 00:02:59,440 So then we need to annotate those images. 39 00:02:59,440 --> 00:03:03,310 And then I need to train the object detection model on this data set. 40 00:03:03,310 --> 00:03:07,000 So this is of course very much time consuming okay. 41 00:03:07,000 --> 00:03:12,970 So uh due to this limitation, like we have to, uh, collect the data set and data data set and train 42 00:03:12,970 --> 00:03:14,200 the object detection model. 43 00:03:14,200 --> 00:03:18,370 So this is what we need to do in case of traditional object detection model. 44 00:03:18,370 --> 00:03:23,680 So due to this limitations the researchers began to develop open vocabulary models. 45 00:03:23,680 --> 00:03:30,940 So the traditional object detectors are object detector detectors are basically fixed vocabulary detection 46 00:03:30,940 --> 00:03:31,570 models. 47 00:03:31,570 --> 00:03:37,810 So the researchers that have done tried to develop some open vocabulary detection models like YOLO world 48 00:03:37,810 --> 00:03:40,060 is an open vocabulary detection model. 49 00:03:40,060 --> 00:03:43,840 Ground in Reno is an open vocabulary detection model. 50 00:03:43,840 --> 00:03:47,770 So due to this limitations like so much time consuming process. 51 00:03:47,770 --> 00:03:53,590 So the due to this limitations, researchers begin to develop open vocabulary detection models. 52 00:03:53,590 --> 00:03:59,230 A few months back, uh, like a few months back, Ronaldinho was introduced, which is an open vocabulary 53 00:03:59,230 --> 00:03:59,980 detection model. 54 00:03:59,980 --> 00:04:02,590 Or you can say zero shot detection model. 55 00:04:02,590 --> 00:04:05,080 Why we say it a zero shot object detection model. 56 00:04:05,080 --> 00:04:09,520 Because, uh, we do not require to train the model to do object detection. 57 00:04:10,560 --> 00:04:17,490 So in zero short object detection model or in open vocabulary detection object detection models, all 58 00:04:17,490 --> 00:04:22,440 we have to do is prompt the model by specifying the list of classes that you are looking for. 59 00:04:22,440 --> 00:04:25,980 So in case of hello world, we will see uh, in next lecture. 60 00:04:25,980 --> 00:04:31,740 So in case of YOLO world, what we need to do is we just need to specify the list of classes that we 61 00:04:31,740 --> 00:04:33,930 want to detect in an image or video. 62 00:04:33,930 --> 00:04:40,080 And we just want to need to pass that list of classes into the input prompt of the YOLO world model. 63 00:04:40,080 --> 00:04:40,590 Okay. 64 00:04:41,100 --> 00:04:47,370 And we don't need to train the model on that specific classes that you want to detect in image or video. 65 00:04:48,660 --> 00:04:53,880 So this is what it, uh, it's we're doing, uh, open vocabulary detection model, which include the 66 00:04:53,880 --> 00:04:55,230 underworld and growling dino. 67 00:04:55,230 --> 00:05:01,410 We just need to pass the list of classes that we want to detect in image or video, uh, as a prompt 68 00:05:01,410 --> 00:05:03,750 to the underworld or bounding dino model. 69 00:05:03,750 --> 00:05:06,690 But YOLO world outperforms the grounding dino model. 70 00:05:06,690 --> 00:05:11,130 There is a some limitations or you can say downsides of grounding Dino model. 71 00:05:11,130 --> 00:05:16,620 So the downside of downside of grounding dino model is its speed, so that the grounding dino model 72 00:05:16,620 --> 00:05:19,860 takes around one second to process a single image. 73 00:05:20,160 --> 00:05:25,500 Uh, which is good enough if you don't care about the amount of latency, but pretty slow if you are 74 00:05:25,500 --> 00:05:29,760 thinking about processing a live video streams using grounding Dino model. 75 00:05:31,310 --> 00:05:34,670 This reason is that the grounding model and other zeros. 76 00:05:35,060 --> 00:05:41,780 Uh uh, zero shot object detector models, except for the world, use heavy transformer based architecture 77 00:05:41,780 --> 00:05:46,610 and require simultaneous processing of text and images during the inference. 78 00:05:46,610 --> 00:05:54,110 And that slows down the processing and it increases the inference time or it increases the latency. 79 00:05:54,350 --> 00:05:54,920 Okay. 80 00:05:55,220 --> 00:05:59,720 And here comes the YOLO world, uh, which is a zero shot object detector model. 81 00:05:59,720 --> 00:06:02,690 And it is equally accurate, unlike Grounding Dino. 82 00:06:02,690 --> 00:06:08,720 And it is 20 times faster than grounded Dino and other, uh, zero shot models. 83 00:06:08,720 --> 00:06:13,910 And it is five times smaller than its predecessors, like Grounding Dino. 84 00:06:14,420 --> 00:06:20,540 So the YOLO world model outperforms other zero shot object detection models like it is 20 times faster, 85 00:06:20,960 --> 00:06:23,450 five times smaller, and it is equally accurate. 86 00:06:25,290 --> 00:06:26,370 So going ahead. 87 00:06:26,370 --> 00:06:27,750 So what is YOLO world? 88 00:06:27,870 --> 00:06:34,320 So YOLO world is introduced in the research paper, uh, title YOLO World Real Time Open Vocabulary 89 00:06:34,320 --> 00:06:40,050 Object Detection, which shows a significant advancement in the field of open vocabulary object detection 90 00:06:40,050 --> 00:06:45,810 by demonstrating that lightweight detector, uh, such as these from the YOLO series. 91 00:06:45,810 --> 00:06:53,760 So like the YOLO world model, use faster CNN based YOLO architecture like in YOLO E8, and it achieves 92 00:06:53,760 --> 00:06:55,350 a strong vocabulary performance. 93 00:06:55,350 --> 00:06:55,920 Like it? 94 00:06:55,920 --> 00:07:01,290 Uh, it is equally accurate than the other zero shot object detection world object detector models. 95 00:07:01,320 --> 00:07:08,130 Okay, so YOLO world basically introduces the prompt then detect paradigm which is a novel approach. 96 00:07:08,130 --> 00:07:14,340 So YOLO world a novel approach by the name prompt that then detect paradigm is being introduced that 97 00:07:14,340 --> 00:07:20,880 avoids the need for real time text encoding or like a property of other zero shot object detector, 98 00:07:21,120 --> 00:07:23,220 uh, object detector models like grounding. 99 00:07:23,640 --> 00:07:28,740 So in ground, uh, in YOLO world, we don't perform real time text encoding. 100 00:07:28,740 --> 00:07:31,530 That basically reduces the speed of the model. 101 00:07:31,530 --> 00:07:37,080 And this real time text encoding is being performed in grounding Reno and other zero shot object detector 102 00:07:37,080 --> 00:07:38,370 models as well. 103 00:07:39,790 --> 00:07:43,150 So the yellow world provides comes with three different models. 104 00:07:43,150 --> 00:07:49,690 You have the world small, which has 313 million parameters, and when Reparameterized goes to 77 million, 105 00:07:49,690 --> 00:07:55,810 and the yellow world medium model has 29 million parameters, and when it is reparameterized, it goes 106 00:07:55,810 --> 00:07:56,680 to 92 million. 107 00:07:56,680 --> 00:08:00,190 And yellow world large model has 48 million parameters. 108 00:08:00,190 --> 00:08:04,810 And when it is reparameterized, it is close to 110 million parameters. 109 00:08:05,140 --> 00:08:11,440 So the yellow world team benchmarked the model on Elvis data set and measured the performance on a V 110 00:08:11,440 --> 00:08:19,120 and V 100 GPU without any performance acceleration mechanism like quantization or tensor RT. 111 00:08:19,720 --> 00:08:26,710 So according to the paper, Yellow World uh reached uh between 35.4 and mean average precision with 112 00:08:26,710 --> 00:08:37,180 52.0 on the large model, uh, and 26.2 mean average precision with 74.1 for the large version. 113 00:08:37,180 --> 00:08:43,240 So this performance is quite, uh, low if we compare with other object detector models. 114 00:08:43,240 --> 00:08:48,550 But this performance is quite compatible if we compare it with other zero shot object detector models. 115 00:08:48,550 --> 00:08:53,680 But if you compare with like state of the art object detector models like YOLO, yolo, v9, this performance 116 00:08:53,680 --> 00:09:00,040 is low, but it is, uh, equally accurate when we compare it with grounding dino Glip like these models. 117 00:09:01,110 --> 00:09:02,790 So here comes the conclusion. 118 00:09:03,270 --> 00:09:11,340 So, is your world a golden solution or is that golden ocean the model that ends training on the custom 119 00:09:11,340 --> 00:09:11,820 dataset? 120 00:09:11,820 --> 00:09:14,910 So do I consider a YOLO world a golden solution? 121 00:09:15,030 --> 00:09:19,950 Like it ends training on the custom dataset, you just need to pass the name of the classes that you 122 00:09:19,950 --> 00:09:26,040 want to detect in image or video in the uh, as a prompt, uh, to the YOLO world model. 123 00:09:26,040 --> 00:09:26,910 So. 124 00:09:27,930 --> 00:09:33,300 But I don't think so that your world is a golden solution, because there are still cases where I would 125 00:09:33,300 --> 00:09:35,940 choose the model trained on a custom data set. 126 00:09:36,270 --> 00:09:41,670 Uh, like the Yolo v eight or YOLO v nine model trained on the custom data set over zero shot object 127 00:09:41,670 --> 00:09:43,140 detector like YOLO world. 128 00:09:44,270 --> 00:09:46,640 Because there is one issue which is latency. 129 00:09:46,640 --> 00:09:52,730 Like the Hello World takes around, uh, hello world takes more time than other state of the art object 130 00:09:52,730 --> 00:09:59,510 detector, uh, to process an image or the latency is a much higher than the other state of the art 131 00:09:59,510 --> 00:10:00,800 object detection models. 132 00:10:01,340 --> 00:10:07,670 So although YOLO world is faster than Reno, but it is slow as compared to other object detection models 133 00:10:07,670 --> 00:10:09,260 like V8 or YOLO Nye. 134 00:10:11,080 --> 00:10:16,990 Okay, so if we required faster processing and have limited limited computational resources, for example, 135 00:10:16,990 --> 00:10:23,770 if I'm working on an Nvidia T4 GPU, like I have limited, uh, the resources like limited GPU requirements 136 00:10:23,770 --> 00:10:29,860 and uh, I require fast processing in real time, then I will be using the traditional object detectors 137 00:10:29,860 --> 00:10:37,000 like YOLO v9, Yolo V8 plus Yolo V8 is less accurate and less reliable as compared to other object detectors 138 00:10:37,000 --> 00:10:42,370 like YOLO seven, Yolo V8 or YOLO v nine when trained or custom data set. 139 00:10:42,370 --> 00:10:49,390 So in short, YOLO world is is an important step in making open vocabulary object detection models faster, 140 00:10:49,390 --> 00:10:55,930 cheaper and widely available, making nearly the same accuracy like other its predecessors like Ground 141 00:10:55,930 --> 00:10:56,290 and Reno. 142 00:10:56,290 --> 00:11:02,800 But and YOLO we will avoid is 20 times faster and five, uh, times smaller than the leading zero shot 143 00:11:03,340 --> 00:11:04,600 detectors models. 144 00:11:04,600 --> 00:11:06,010 So that's all from hysteria. 145 00:11:06,040 --> 00:11:06,340 Thank you.