# Detectnet\_v2(resnet50) low accuracy on 2 class dataset **URL:** \ **Category:** TAO Toolkit\ **Created:** [January 30, 2023, 9:36am UTC](https://forums.developer.nvidia.com/t/detectnet-v2-resnet50-low-accuracy-on-2-class-dataset/241013 "2023-01-30T09:36:44Z")\ **Posts on this page:** 6\ **Page:** 2
**Author:** ![Morganh](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/morganh/32/9748_2.png) [@Morganh](https://forums.developer.nvidia.com/u/Morganh)\ **Post date:** [February 10, 2023, 4:16pm UTC](https://forums.developer.nvidia.com/t/detectnet-v2-resnet50-low-accuracy-on-2-class-dataset/241013/22 "2023-02-10T16:16:57Z")
> [@pddarrell](#): > > Does the ubiquitous result of “0” mean something is not set up correctly? - zero is difficult to achieve, isn’t it? I am afraid yes. Please check 1. Are the labels correct? Especially for the “healthy” class. 2. More, how many images for these two classes separately? 3. Set to the same initial\_weight: 10.0 weight\_target: 10.0 4. Can you share latest training spec? ---
**Author:** ![pddarrell](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@pddarrell](https://forums.developer.nvidia.com/u/pddarrell)\ **Post date:** [February 10, 2023, 5:18pm UTC](https://forums.developer.nvidia.com/t/detectnet-v2-resnet50-low-accuracy-on-2-class-dataset/241013/23 "2023-02-10T17:18:15Z")
> [@Morganh](#): > > Are the labels correct? Especially for the “healthy” class. “healthy” image and label files have a “0” as the first digit: They start at “000001” and are not entirely consecutive e.g: ![000016](https://global.discourse-cdn.com/nvidia/original/3X/1/3/132cb4cd7154fad964da935ee031e0d0be9047fd.jpeg) [000016.txt](https://forums.developer.nvidia.com/uploads/short-url/bin2LaaCnOikhhTcWyGGVcJ8xqh.txt) (76 Bytes) “damage” image and label files have a “1” as the first digit: ![286](https://global.discourse-cdn.com/nvidia/original/3X/5/8/583d5d9af1c821ac72ed5fa705c0fad09a9529c0.jpeg) [100268.txt](https://forums.developer.nvidia.com/uploads/short-url/5ezVPs6qajEctjhCYdE93NojoOn.txt) (74 Bytes) I believe the labels are 100% correct. I hand-labelled them and I have checked them very carefully. > Blockquote ![Screenshot from 2023-02-10 16-48-49](https://global.discourse-cdn.com/nvidia/original/3X/d/b/dba8ac9c53761b8ca38e2be123a51bd044dddd63.png) In train/val: 531 “healthy”, 531 “damage” In test: 10 “healthy”, 10 “damage” > Blockquote > currently the settings in “cost\_function\_config” are: cost\_function\_config { target\_classes { name: “damage” class\_weight: 8.0 coverage\_foreground\_weight: 0.0500000007451 objectives { name: “cov” initial\_weight: 1.0 weight\_target: 1.0 } objectives { name: “bbox” initial\_weight: 10.0 weight\_target: 1.0 } } target\_classes { name: “healthy” class\_weight: 1.0 coverage\_foreground\_weight: 0.0500000007451 objectives { name: “cov” initial\_weight: 1.0 weight\_target: 1.0 } objectives { name: “bbox” initial\_weight: 10.0 weight\_target: 10.0 } } To be 100% sure: do you want me to reset the initial\_weight and weight\_target of both classes to 10.0? > Blockquote > [detectnet\_v2\_train\_resnet50\_kitti.txt](https://forums.developer.nvidia.com/uploads/short-url/f5AMoXkv9iOeJg4GclLRSG5RYXp.txt) (4.2 KB) > This is the latest training spec before I reset the initial\_weight and weight\_target of both classes to 10.0. ---
**Author:** ![Morganh](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/morganh/32/9748_2.png) [@Morganh](https://forums.developer.nvidia.com/u/Morganh)\ **Post date:** [February 11, 2023, 4:17am UTC](https://forums.developer.nvidia.com/t/detectnet-v2-resnet50-low-accuracy-on-2-class-dataset/241013/24 "2023-02-11T04:17:48Z")
For healthy class, from the label file, ```auto healthy 0.00 0 0.00 0.0 0.0 3648.0 1417.0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 ``` It is 1920x746 image, why the bbox is (0.0 0.0 3648.0 1417.0) ? ---
**Author:** ![pddarrell](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@pddarrell](https://forums.developer.nvidia.com/u/pddarrell)\ **Post date:** [February 11, 2023, 12:36pm UTC](https://forums.developer.nvidia.com/t/detectnet-v2-resnet50-low-accuracy-on-2-class-dataset/241013/25 "2023-02-11T12:36:52Z")
> [@Morganh](#): > > It is 1920x746 image, why the bbox is (0.0 0.0 3648.0 1417.0) ? ![Screenshot from 2023-02-11 12-31-19](https://global.discourse-cdn.com/nvidia/original/3X/3/a/3a628e8aa4ab7c20b1c561a47df4858f448a68c6.png) image size of 000016.jpg is 3648 x 1417. I guess the forum formatting has resized the image that you are seeing. The label is correct. ---
**Author:** ![pddarrell](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@pddarrell](https://forums.developer.nvidia.com/u/pddarrell)\ **Post date:** [February 12, 2023, 7:24pm UTC](https://forums.developer.nvidia.com/t/detectnet-v2-resnet50-low-accuracy-on-2-class-dataset/241013/26 "2023-02-12T19:24:20Z")
> [@pddarrell](#): > > Does the notebook allow the confidence threshold to be adjusted? Maybe it is a confidence issue? Problem solved. I had the coverage\_threshold for “healthy” in postprocessing\_config set too high. Interestingly, when I lowered it the first time the average precision of “damage” improved and “healthy” stayed at “0”. It encouraged me to lower it further and, finally, I got a non-zero value for “healthy” too. ![Screenshot from 2023-02-12 19-22-28](https://global.discourse-cdn.com/nvidia/original/3X/7/2/722cff5428539575a4cc7d0e0d7b7065bb216abc.png) These experiments are only run for 10 epochs, so I am hopeful that further improvements are possible at ~ 120 epochs. Basically I had not fully appreciated what “mean average precision” was doing and what it was affected by. I found this blog helpful: [https://machinethink.net/blog/object-detection](https://machinethink.net/blog/object-detection) Thank you for being a great help. ---
**Author:** ![system](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/system/32/68080_2.png) [@system](https://forums.developer.nvidia.com/u/system)\ **Post date:** [February 26, 2023, 7:24pm UTC](https://forums.developer.nvidia.com/t/detectnet-v2-resnet50-low-accuracy-on-2-class-dataset/241013/27 "2023-02-26T19:24:57Z")
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed. [Previous page](https://forums.developer.nvidia.com/t/detectnet-v2-resnet50-low-accuracy-on-2-class-dataset/241013.md?page=1)