Recherche avancée

Sur d’autres sites (9609)

• How to Choose the Optimal Multi-Touch Attribution Model for Your Organisation

  13 mars 2023, par Erin — Analytics Tips

  If you struggle to connect the dots on your customer journeys, you are researching the correct solution. 

  Multi-channel attribution models allow you to better understand the users’ paths to conversion and identify key channels and marketing assets that assist them.

  That said, each attribution model has inherent limitations, which make the selection process even harder.

  This guide explains how to choose the optimal multi-touch attribution model. We cover the pros and cons of popular attribution models, main evaluation criteria and how-to instructions for model implementation. 

  Pros and Cons of Different Attribution Models 

  

  First Interaction 

  First Interaction attribution model (also known as first touch) assigns full credit to the conversion to the first channel, which brought in a lead. However, it doesn’t report other interactions the visitor had before converting.

  Marketers, who are primarily focused on demand generation and user acquisition, find the first touch attribution model useful to evaluate and optimise top-of-the-funnel (ToFU). 

  Pros 

  • Reflects the start of the customer journey
  • Shows channels that bring in the best-qualified leads 
  • Helps track brand awareness campaigns

  Cons 

  • Ignores the impact of later interactions at the middle and bottom of the funnel 
  • Doesn’t provide a full picture of users’ decision-making process 

  Last Interaction 

  Last Interaction attribution model (also known as last touch) shifts the entire credit allocation to the last channel before conversion. But it doesn’t account for the contribution of all other channels. 

  If your focus is conversion optimization, the last-touch model helps you determine which channels, assets or campaigns seal the deal for the prospect. 

  Pros 

  • Reports bottom-of-the-funnel events
  • Requires minimal data and configurations 
  • Helps estimate cost-per-lead or cost-per-acquisition

  Cons 

  • No visibility into assisted conversions and prior visitor interactions 
  • Overemphasise the importance of the last channel (which can often be direct traffic) 

  Last Non-Direct Interaction 

  Last Non-Direct attribution excludes direct traffic from the calculation and assigns the full conversion credit to the preceding channel. For example, a paid ad will receive 100% of credit for conversion if a visitor goes directly to your website to buy a product. 

  Last Non-Direct attribution provides greater clarity into the bottom-of-the-funnel (BoFU). events. Yet, it still under-reports the role other channels played in conversion. 

  Pros 

  • Improved channel visibility, compared to Last-Touch 
  • Avoids over-valuing direct visits
  • Reports on lead-generation efforts

  Cons 

  • Doesn’t work for account-based marketing (ABM) 
  • Devalues the quality over quantity of leads 

  Linear Model

  Linear attribution model assigns equal credit for a conversion to all tracked touchpoints, regardless of their impact on the visitor’s decision to convert.

  It helps you understand the full conversion path. But this model doesn’t distinguish between the importance of lead generation activities versus nurturing touches.

  Pros 

  • Focuses on all touch points associated with a conversion 
  • Reflects more steps in the customer journey 
  • Helps analyse longer sales cycles

  Cons 

  • Doesn’t accurately reflect the varying roles of each touchpoint 
  • Can dilute the credit if too many touchpoints are involved 

  Time Decay Model 

  Time decay models assumes that the closer a touchpoint is to the conversion, the greater its influence. Pre-conversion touchpoints get the highest credit, while the first ones are ranked lower (5%-5%-10%-15%-25%-30%).

  This model better reflects real-life customer journeys. However, it devalues the impact of brand awareness and demand-generation campaigns. 

  Pros 

  • Helps track longer sales cycles and reports on each touchpoint involved 
  • Allows customising the half-life of decay to improve reporting 
  • Promotes conversion optimization at BoFu stages

  Cons 

  • Can prompt marketers to curtail ToFU spending, which would translate to fewer qualified leads at lower stages
  • Doesn’t reflect highly-influential events at earlier stages (e.g., a product demo request or free account registration, which didn’t immediately lead to conversion)

  Position-Based Model 

  Position-Based attribution model (also known as the U-shaped model) allocates the biggest credit to the first and the last interaction (40% each). Then distributes the remaining 20% across other touches. 

  For many marketers, that’s the preferred multi-touch attribution model as it allows optimising both ToFU and BoFU channels. 

  Pros 

  • Helps establish the main channels for lead generation and conversion
  • Adds extra layers of visibility, compared to first- and last-touch attribution models 
  • Promotes budget allocation toward the most strategic touchpoints

  Cons 

  • Diminishes the importance of lead nurturing activities as more credit gets assigned to demand-gen and conversion-generation channels
  • Limited flexibility since it always assigns a fixed amount of credit to the first and last touchpoints, and the remaining credit is divided evenly among the other touchpoints

  How to Choose the Right Multi-Touch Attribution Model For Your Business 

  If you’re deciding which attribution model is best for your business, prepare for a heated discussion. Each one has its trade-offs as it emphasises or devalues the role of different channels and marketing activities.

  To reach a consensus, the best strategy is to evaluate each model against three criteria : Your marketing objectives, sales cycle length and data availability. 

  Marketing Objectives 

  Businesses generate revenue in many ways : Through direct sales, subscriptions, referral fees, licensing agreements, one-off or retainer services. Or any combination of these activities. 

  In each case, your marketing strategy will look different. For example, SaaS and direct-to-consumer (DTC) eCommerce brands have to maximise both demand generation and conversion rates. In contrast, a B2B cybersecurity consulting firm is more interested in attracting qualified leads (as opposed to any type of traffic) and progressively nurturing them towards a big-ticket purchase. 

  When selecting a multi-touch attribution model, prioritise your objectives first. Create a simple scoreboard, where your team ranks various channels and campaign types you rely on to close sales. 

  Alternatively, you can survey your customers to learn how they first heard about your company and what eventually triggered their conversion. Having data from both sides can help you cross-validate your assumptions and eliminate some biases. 

  Then consider which model would best reflect the role and importance of different channels in your sales cycle. Speaking of which….

  Sales Cycle Length 

  As shoppers, we spend less time deciding on a new toothpaste brand versus contemplating a new IT system purchase. Factors like industry, business model (B2C, DTC, B2B, B2BC), and deal size determine the average cycle length in your industry. 

  Statistically, low-ticket B2C sales can happen within just several interactions. The average B2B decision-making process can have over 15 steps, spread over several months. 

  That’s why not all multi-touch attribution models work equally well for each business. Time-decay suits better B2B companies, while B2C usually go for position-based or linear attribution. 

  Data Availability 

  Businesses struggle with multi-touch attribution model implementation due to incomplete analytics data. 

  Our web analytics tool captures more data than Google Analytics. That’s because we rely on a privacy-focused tracking mechanism, which allows you to collect analytics without showing a cookie consent banner in markets outside of Germany and the UK. 

  Cookie consent banners are mandatory with Google Analytics. Yet, almost 40% of global consumers reject it. This results in gaps in your analytics and subsequent inconsistencies in multi-touch attribution reports. With Matomo, you can compliantly collect more data for accurate reporting. 

  Some companies also struggle to connect collected insights to individual shoppers. With Matomo, you can cross-attribute users across browning sessions, using our visitors’ tracking feature. 

  When you already know a user’s identifier (e.g., full name or email address), you can track their on-site behaviours over time to better understand how they interact with your content and complete their purchases. Quick disclaimer, though, visitors’ tracking may not be considered compliant with certain data privacy laws. Please consult with a local authority if you have doubts. 

  How to Implement Multi-Touch Attribution

  Multi-touch attribution modelling implementation is like a “seek and find” game. You have to identify all significant touchpoints in your customers’ journeys. And sometimes also brainstorm new ways to uncover the missing parts. Then figure out the best way to track users’ actions at those stages (aka do conversion and events tracking). 

  Here’s a step-by-step walkthrough to help you get started. 

  Select a Multi-Touch Attribution Tool 

  The global marketing attribution software is worth $3.1 billion. Meaning there are plenty of tools, differing in terms of accuracy, sophistication and price.

  To make the right call prioritise five factors :

  • Available models : Look for a solution that offers multiple options and allows you to experiment with different modelling techniques or develop custom models. 
  • Implementation complexity : Some providers offer advanced data modelling tools for creating custom multi-touch attribution models, but offer few out-of-the-box modelling options. 
  • Accuracy : Check if the shortlisted tool collects the type of data you need. Prioritise providers who are less dependent on third-party cookies and allow you to identify repeat users. 
  • Your marketing stack : Some marketing attribution tools come with useful add-ons such as tag manager, heatmaps, form analytics, user session recordings and A/B testing tools. This means you can collect more data for multi-channel modelling with them instead of investing in extra software. 
  • Compliance : Ensure that the selected multi-attribution analytics software wouldn’t put you at risk of GDPR non-compliance when it comes to user privacy and consent to tracking/analysis. 

  Finally, evaluate the adoption costs. Free multi-channel analytics tools come with data quality and consistency trade-offs. Premium attribution tools may have “hidden” licensing costs and bill you for extra data integrations. 

  Look for a tool that offers a good price-to-value ratio (i.e., one that offers extra perks for a transparent price). 

  Set Up Proper Data Collection 

  Multi-touch attribution requires ample user data. To collect the right type of insights you need to set up : 

  • Website analytics : Ensure that you have all tracking codes installed (and working correctly !) to capture pageviews, on-site actions, referral sources and other data points around what users do on page. 
  • Tags : Add tracking parameters to monitor different referral channels (e.g., “facebook”), campaign types (e.g., ”final-sale”), and creative assets (e.g., “banner-1”). Tags help you get a clearer picture of different touchpoints. 
  • Integrations : To better identify on-site users and track their actions, you can also populate your attribution tool with data from your other tools – CRM system, A/B testing app, etc. 

  Finally, think about the ideal lookback window — a bounded time frame you’ll use to calculate conversions. For example, Matomo has a default windows of 7, 30 or 90 days. But you can configure a custom period to better reflect your average sales cycle. For instance, if you’re selling makeup, a shorter window could yield better results. But if you’re selling CRM software for the manufacturing industry, consider extending it.

  Configure Goals and Events 

  Goals indicate your main marketing objectives — more traffic, conversions and sales. In web analytics tools, you can measure these by tracking specific user behaviours. 

  For example : If your goal is lead generation, you can track :

  • Newsletter sign ups 
  • Product demo requests 
  • Gated content downloads 
  • Free trial account registration 
  • Contact form submission 
  • On-site call bookings 

  In each case, you can set up a unique tag to monitor these types of requests. Then analyse conversion rates — the percentage of users who have successfully completed the action. 

  To collect sufficient data for multi-channel attribution modelling, set up Goal Tracking for different types of touchpoints (MoFU & BoFU) and asset types (contact forms, downloadable assets, etc). 

  Your next task is to figure out how users interact with different on-site assets. That’s when Event Tracking comes in handy. 

  Event Tracking reports notify you about specific actions users take on your website. With Matomo Event Tracking, you can monitor where people click on your website, on which pages they click newsletter subscription links, or when they try to interact with static content elements (e.g., a non-clickable banner). 

  Using in-depth user behavioural reports, you can better understand which assets play a key role in the average customer journey. Using this data, you can localise “leaks” in your sales funnel and fix them to increase conversion rates.

  Test and Validated the Selected Model 

  A common challenge of multi-channel attribution modelling is determining the correct correlation and causality between exposure to touchpoints and purchases. 

  For example, a user who bought a discounted product from a Facebook ad would act differently than someone who purchased a full-priced product via a newsletter link. Their rate of pre- and post-sales exposure will also differ a lot — and your attribution model may not always accurately capture that. 

  That’s why you have to continuously test and tweak the selected model type. The best approach for that is lift analysis. 

  Lift analysis means comparing how your key metrics (e.g., revenue or conversion rates) change among users who were exposed to a certain campaign versus a control group. 

  In the case of multi-touch attribution modelling, you have to monitor how your metrics change after you’ve acted on the model recommendations (e.g., invested more in a well-performing referral channel or tried a new brand awareness Twitter ad). Compare the before and after ROI. If you see a positive dynamic, your model works great. 

  The downside of this approach is that you have to invest a lot upfront. But if your goal is to create a trustworthy attribution model, the best way to validate is to act on its suggestions and then test them against past results. 

  Conclusion

  A multi-touch attribution model helps you measure the impact of different channels, campaign types, and marketing assets on metrics that matter — conversion rate, sales volumes and ROI. 

  Using this data, you can invest budgets into the best-performing channels and confidently experiment with new campaign types. 

  As a Matomo user, you also get to do so without breaching customers’ privacy or compromising on analytics accuracy.

  Start using accurate multi-channel attribution in Matomo. Get your free 21-day trial now. No credit card required.

• What's the most desireable way to capture system display and audio in the form of individual encoded audio and video packets in go (language) ? [closed]

  11 janvier 2023, par Tiger Yang

  Question (read the context below first) :

  


  For those of you familiar with the capabilities of go, Is there a better way to go about all this ? Since ffmpeg is so ubiquitous, I'm sure it's been optomized to perfection, but what's the best way to capture system display and audio in the form of individual encoded audio and video packets in go (language), so that they can be then sent via webtransport-go ? I wish for it to prioritize efficiency and low latency, and ideally capture and encode the framebuffer directly like ffmpeg does.

  


  Thanks ! I have many other questions about this, but I think it's best to ask as I go.

  


  Context and what I've done so far :

  


  I'm writing a remote desktop software for my personal use because of grievances with current solutions out there. At the moment, it consists of a web app that uses the webtransport API to send input datagrams and receive AV packets on two dedicated unidirectional streams, and the webcodecs API to decode these packets. On the serverside, I originally planned to use python with the aioquic library as a webtransport server. Upon connection and authentication, the server would start ffmpeg as a subprocess with this command :

  


  ffmpeg -init_hw_device d3d11va -filter_complex ddagrab=video_size=1920x1080:framerate=60 -vcodec hevc_nvenc -tune ll -preset p7 -spatial_aq 1 -temporal_aq 1 -forced-idr 1 -rc cbr -b:v 400K -no-scenecut 1 -g 216000 -f hevc -

  


  What I really appreciate about this is that it uses windows' desktop duplication API to copy the framebuffer of my GPU and hand that directly to the on-die hardware encoder with zero round trips to the CPU. I think it's about as efficient and elegant a solution as I can manage. It then outputs the encoded stream to the stdout, which python can read and send to the client.

  


  As for the audio, there is another ffmpeg instance :

  


  ffmpeg -f dshow -channels 2 -sample_rate 48000 -sample_size 16 -audio_buffer_size 15 -i audio="RD Audio (High Definition Audio Device)" -acodec libopus -vbr on -application audio -mapping_family 0 -apply_phase_inv true -b:a 25K -fec false -packet_loss 0 -map 0 -f data -

  


  which listens to a physical loopback interface, which is literally just a short wire bridging the front panel headphone and microphone jacks (I'm aware of the quality loss of converting to analog and back, but the audio is then crushed down to 25kbps so it's fine) ()

  


  Unfortunately, aioquic was not easy to work with IMO, and I found webtransport-go https://github.com/adriancable/webtransport-go, which was a hell of a lot better in both simplicity and documentation. However, now I'm dealing with a whole new language, and I wanna ask : (above)

  


  EDIT : Here's the code for my server so far :

  


  


  


  package main

import (
    "bytes"
    "context"
    "fmt"
    "log"
    "net/http"
    "os/exec"
    "time"

    "github.com/adriancable/webtransport-go"
)

func warn(str string) {
    fmt.Printf("\n===== WARNING ===================================================================================================\n   %s\n=================================================================================================================\n", str)
}

func main() {

    password := []byte("abc")

    videoString := []string{
        "ffmpeg",
        "-init_hw_device", "d3d11va",
        "-filter_complex", "ddagrab=video_size=1920x1080:framerate=60",
        "-vcodec", "hevc_nvenc",
        "-tune", "ll",
        "-preset", "p7",
        "-spatial_aq", "1",
        "-temporal_aq", "1",
        "-forced-idr", "1",
        "-rc", "cbr",
        "-b:v", "500K",
        "-no-scenecut", "1",
        "-g", "216000",
        "-f", "hevc", "-",
    }

    audioString := []string{
        "ffmpeg",
        "-f", "dshow",
        "-channels", "2",
        "-sample_rate", "48000",
        "-sample_size", "16",
        "-audio_buffer_size", "15",
        "-i", "audio=RD Audio (High Definition Audio Device)",
        "-acodec", "libopus",
        "-mapping_family", "0",
        "-b:a", "25K",
        "-map", "0",
        "-f", "data", "-",
    }

    connected := false

    http.HandleFunc("/", func(writer http.ResponseWriter, request *http.Request) {
        session := request.Body.(*webtransport.Session)

        session.AcceptSession()
        fmt.Println("\nAccepted incoming WebTransport connection.")
        fmt.Println("Awaiting authentication...")

        authData, err := session.ReceiveMessage(session.Context()) // Waits here till first datagram
        if err != nil {                                            // if client closes connection before sending anything
            fmt.Println("\nConnection closed:", err)
            return
        }

        if len(authData) >= 2 && bytes.Equal(authData[2:], password) {
            if connected {
                session.CloseSession()
                warn("Client has authenticated, but a session is already taking place! Connection closed.")
                return
            } else {
                connected = true
                fmt.Println("Client has authenticated!\n")
            }
        } else {
            session.CloseSession()
            warn("Client has failed authentication! Connection closed. (" + string(authData[2:]) + ")")
            return
        }

        videoStream, _ := session.OpenUniStreamSync(session.Context())

        videoCmd := exec.Command(videoString[0], videoString[1:]...)
        go func() {
            videoOut, _ := videoCmd.StdoutPipe()
            videoCmd.Start()

            buffer := make([]byte, 15000)
            for {
                len, err := videoOut.Read(buffer)
                if err != nil {
                    break
                }
                if len > 0 {
                    videoStream.Write(buffer[:len])
                }
            }
        }()

        time.Sleep(50 * time.Millisecond)

        audioStream, err := session.OpenUniStreamSync(session.Context())

        audioCmd := exec.Command(audioString[0], audioString[1:]...)
        go func() {
            audioOut, _ := audioCmd.StdoutPipe()
            audioCmd.Start()

            buffer := make([]byte, 15000)
            for {
                len, err := audioOut.Read(buffer)
                if err != nil {
                    break
                }
                if len > 0 {
                    audioStream.Write(buffer[:len])
                }
            }
        }()

        for {
            data, err := session.ReceiveMessage(session.Context())
            if err != nil {
                videoCmd.Process.Kill()
                audioCmd.Process.Kill()

                connected = false

                fmt.Println("\nConnection closed:", err)
                break
            }

            if len(data) == 0 {

            } else if data[0] == byte(0) {
                fmt.Printf("Received mouse datagram: %s\n", data)
            }
        }

    })

    server := &webtransport.Server{
        ListenAddr: ":1024",
        TLSCert:    webtransport.CertFile{Path: "SSL/fullchain.pem"},
        TLSKey:     webtransport.CertFile{Path: "SSL/privkey.pem"},
        QuicConfig: &webtransport.QuicConfig{
            KeepAlive:      false,
            MaxIdleTimeout: 3 * time.Second,
        },
    }

    fmt.Println("Launching WebTransport server at", server.ListenAddr)
    ctx, cancel := context.WithCancel(context.Background())
    if err := server.Run(ctx); err != nil {
        log.Fatal(err)
        cancel()
    }

}

  


  


  



• Using FFmpeg to stitch together H.264 videos and variably-spaced JPEG pictures ; dealing with ffmpeg warnings

  19 octobre 2022, par LB2

  Context

  


  I have a process flow that may output either H.264 Annex B streams, variably-spaced JPEGs, or a mixture of two. By variably-spaced I mean where elapsed time between any two adjacent JPEGs may (and likely to be) different from any other two adjacent JPEGs. So an example of possible inputs are :

  


  

  1. stream1.h264

  


  2. {Set of JPEGs}

  


  3. stream1.h264 + stream2.h264

  


  4. stream1.h264 + {Set of JPEGs}

  


  5. stream1.h264 + {Set of JPEGs} + stream2.h264

  


  6. stream1.h264 + {Set of JPEGs} + stream2.h264 + {Set of JPEGs} + ...

  


  7. stream1.h264 + stream2.h264 + {Set of JPEGs} + ...

  


  


  The output needs to be a single stitched (i.e. concatenated) output in MPEG-4 container.

  


  Requirements : No re-encoding or transcoding of existing video compression (One time conversion of JPEG sets to video format is OKay).

  


  Solution Prototype

  


  To prototype the solution I have found that ffmpeg has concat demuxer that would let me specify an ordered sequence of inputs that ffmpeg would then concatenate together, but all inputs must be of the same format. So, to meet that requirement, I :

  


  

  1. Convert every JPEG set to an .mp4 using concat (and using delay # directive to specify time-spacing between each JPEG)

  


  2. Convert every .h264 to .mp4 using -c copy to avoid transcoding.

  


  3. Stitch all generated interim .mp4 files into the single final .mp4 using -f concat and -c copy.

  


  


  Here's the bash script, in parts, that performs the above :

  


  


  1. Ignore the curl comment ; that's from originally generating a 100 jpeg images with numbers and these are simply saved locally. What the loop does is it generates concat input file with file sequence#.jpeg directives and duration # directive where each successive JPEG delay is incremented by 0.1 seconds (0.1 between first and second, 0.2 b/w 2nd and 3rd, 0.3 b/w 3rd and 4th, and so on). Then it runs ffmpeg command to convert the set of JPEGs to .mp4 interim file.

     


     echo "ffconcat version 1.0" >ffconcat-jpeg.txt
echo >>ffconcat-jpeg.txt

for i in {1..100}
do
    echo "file $i.jpeg" >>ffconcat-jpeg.txt
    d=$(echo "$i" | awk '{printf "%f", $1 / 10}')
    # d=$(echo "scale=2; $i/10" | bc)
    echo "duration $d" >>ffconcat-jpeg.txt
    echo "" >>ffconcat-jpeg.txt
    # curl -o "$i.jpeg" "https://math.tools/equation/get_equaimages?equation=$i&fontsize=256"
done

ffmpeg \
    -hide_banner \
    -vsync vfr \
    -f concat \
    -i ffconcat-jpeg.txt \
    -r 30 \
    -video_track_timescale 90000 \
    video-jpeg.mp4


     


  


  2. Convert two streams from .h264 to .mp4 via copy (no transcoding).

     


     ffmpeg \
    -hide_banner \
    -i low-motion-video.h264 \
    -c copy \
    -vsync vfr \
    -video_track_timescale 90000 \
    low-motion-video.mp4

ffmpeg \
    -hide_banner \
    -i full-video.h264 \
    -c copy \
    -video_track_timescale 90000 \
    -vsync vfr \
    full-video.mp4


     


  


  3. Stitch all together by generating another concat directive file.

     


     echo "ffconcat version 1.0" >ffconcat-h264.txt
echo >>ffconcat-h264.txt
echo "file low-motion-video.mp4" >>ffconcat-h264.txt
echo >>ffconcat-h264.txt
echo "file full-video.mp4" >>ffconcat-h264.txt
echo >>ffconcat-h264.txt
echo "file video-jpeg.mp4" >>ffconcat-h264.txt
echo >>ffconcat-h264.txt

ffmpeg \
    -hide_banner \
    -f concat \
    -i ffconcat-h264.txt \
    -pix_fmt yuv420p \
    -c copy \
    -video_track_timescale 90000 \
    -vsync vfr \
    video-out.mp4



     


  


  


  Problem (and attempted troubleshooting)

  


  The above does produce a reasonable output — it plays first video, then plays second video with no timing/rate issues AFAICT, then plays JPEGs with time between each JPEG "frame" growing successively, as expected.

  


  But, the conversion process produces warnings that concern me (for compatibility with players ; or potentially other IRL streams that may result in some issue my prototyping content doesn't make obvious). Initial attempts generated 100s of warnings, but with some arguments added, I reduced it down to just a handful, but this handful is stubborn and nothing I tried would help.

  


  The first conversion of JPEGs to .mp4 goes fine with the following output :

  


  Input #0, concat, from 'ffconcat-jpeg.txt':
  Duration: 00:08:25.00, start: 0.000000, bitrate: 0 kb/s
  Stream #0:0: Video: png, pal8(pc), 176x341 [SAR 3780:3780 DAR 16:31], 25 fps, 25 tbr, 25 tbn, 25 tbc
Stream mapping:
  Stream #0:0 -> #0:0 (png (native) -> h264 (libx264))
Press [q] to stop, [?] for help
[libx264 @ 0x7fe418008e00] using SAR=1/1
[libx264 @ 0x7fe418008e00] using cpu capabilities: MMX2 SSE2Fast SSSE3 SSE4.2
[libx264 @ 0x7fe418008e00] profile High 4:4:4 Predictive, level 1.3, 4:4:4, 8-bit
[libx264 @ 0x7fe418008e00] 264 - core 163 r3060 5db6aa6 - H.264/MPEG-4 AVC codec - Copyleft 2003-2021 - http://www.videolan.org/x264.html - options: cabac=1 ref=3 deblock=1:0:0 analyse=0x3:0x113 me=hex subme=7 psy=1 psy_rd=1.00:0.00 mixed_ref=1 me_range=16 chroma_me=1 trellis=1 8x8dct=1 cqm=0 deadzone=21,11 fast_pskip=1 chroma_qp_offset=4 threads=11 lookahead_threads=1 sliced_threads=0 nr=0 decimate=1 interlaced=0 bluray_compat=0 constrained_intra=0 bframes=3 b_pyramid=2 b_adapt=1 b_bias=0 direct=1 weightb=1 open_gop=0 weightp=2 keyint=250 keyint_min=25 scenecut=40 intra_refresh=0 rc_lookahead=40 rc=crf mbtree=1 crf=23.0 qcomp=0.60 qpmin=0 qpmax=69 qpstep=4 ip_ratio=1.40 aq=1:1.00
Output #0, mp4, to 'video-jpeg.mp4':
  Metadata:
    encoder         : Lavf58.76.100
  Stream #0:0: Video: h264 (avc1 / 0x31637661), yuv444p(tv, progressive), 176x341 [SAR 1:1 DAR 16:31], q=2-31, 30 fps, 90k tbn
    Metadata:
      encoder         : Lavc58.134.100 libx264
    Side data:
      cpb: bitrate max/min/avg: 0/0/0 buffer size: 0 vbv_delay: N/A
frame=  100 fps=0.0 q=-1.0 Lsize=     157kB time=00:07:55.33 bitrate=   2.7kbits/s speed=2.41e+03x    
video:155kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB muxing overhead: 1.800846%
[libx264 @ 0x7fe418008e00] frame I:1     Avg QP:20.88  size:   574
[libx264 @ 0x7fe418008e00] frame P:43    Avg QP:14.96  size:  2005
[libx264 @ 0x7fe418008e00] frame B:56    Avg QP:21.45  size:  1266
[libx264 @ 0x7fe418008e00] consecutive B-frames: 14.0% 24.0% 30.0% 32.0%
[libx264 @ 0x7fe418008e00] mb I  I16..4: 36.4% 55.8%  7.9%
[libx264 @ 0x7fe418008e00] mb P  I16..4:  5.1%  7.5% 11.2%  P16..4:  5.6%  8.1%  4.5%  0.0%  0.0%    skip:57.9%
[libx264 @ 0x7fe418008e00] mb B  I16..4:  2.4%  0.9%  3.9%  B16..8: 16.2%  8.8%  4.6%  direct: 1.2%  skip:62.0%  L0:56.6% L1:38.7% BI: 4.7%
[libx264 @ 0x7fe418008e00] 8x8 transform intra:28.3% inter:3.7%
[libx264 @ 0x7fe418008e00] coded y,u,v intra: 26.5% 0.0% 0.0% inter: 9.8% 0.0% 0.0%
[libx264 @ 0x7fe418008e00] i16 v,h,dc,p: 82% 13%  4%  0%
[libx264 @ 0x7fe418008e00] i8 v,h,dc,ddl,ddr,vr,hd,vl,hu: 20%  8% 71%  1%  0%  0%  0%  0%  0%
[libx264 @ 0x7fe418008e00] i4 v,h,dc,ddl,ddr,vr,hd,vl,hu: 41% 11% 29%  4%  2%  3%  1%  7%  1%
[libx264 @ 0x7fe418008e00] Weighted P-Frames: Y:0.0% UV:0.0%
[libx264 @ 0x7fe418008e00] ref P L0: 44.1%  4.2% 28.4% 23.3%
[libx264 @ 0x7fe418008e00] ref B L0: 56.2% 32.1% 11.6%
[libx264 @ 0x7fe418008e00] ref B L1: 92.4%  7.6%
[libx264 @ 0x7fe418008e00] kb/s:2.50


  


  The conversion of individual streams from .h264 to .mp4 generates two types of warnings each. One is [mp4 @ 0x7faee3040400] Timestamps are unset in a packet for stream 0. This is deprecated and will stop working in the future. Fix your code to set the timestamps properly, and the other is [mp4 @ 0x7faee3040400] pts has no value.

  


  Some posts on SO (can't find my original finds on that now) suggested that it's safe to ignore and comes from H.264 being an elementary stream that supposedly doesn't contain timestamps. It surprises me a bit since I produce that stream using NVENC API and clearly supply timing information for each frame via PIC_PARAMS structure : NV_STRUCT(PIC_PARAMS, pp); ...; pp.inputTimeStamp = _frameIndex++ * (H264_CLOCK_RATE / _params.frameRate);, where #define H264_CLOCK_RATE 9000 and _params.frameRate = 30.

  


  Input #0, h264, from 'low-motion-video.h264':
  Duration: N/A, bitrate: N/A
  Stream #0:0: Video: h264 (High), yuv420p(progressive), 1440x3040 [SAR 1:1 DAR 9:19], 30 fps, 30 tbr, 1200k tbn, 60 tbc
Output #0, mp4, to 'low-motion-video.mp4':
  Metadata:
    encoder         : Lavf58.76.100
  Stream #0:0: Video: h264 (High) (avc1 / 0x31637661), yuv420p(progressive), 1440x3040 [SAR 1:1 DAR 9:19], q=2-31, 30 fps, 30 tbr, 90k tbn, 1200k tbc
Stream mapping:
  Stream #0:0 -> #0:0 (copy)
Press [q] to stop, [?] for help
[mp4 @ 0x7faee3040400] Timestamps are unset in a packet for stream 0. This is deprecated and will stop working in the future. Fix your code to set the timestamps properly
[mp4 @ 0x7faee3040400] pts has no value
[mp4 @ 0x7faee3040400] pts has no value0kB time=-00:00:00.03 bitrate=N/A speed=N/A    
    Last message repeated 17985 times
frame=17987 fps=0.0 q=-1.0 Lsize=   79332kB time=00:09:59.50 bitrate=1084.0kbits/s speed=1.59e+03x    
video:79250kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB muxing overhead: 0.103804%
Input #0, h264, from 'full-video.h264':
  Duration: N/A, bitrate: N/A
  Stream #0:0: Video: h264 (High), yuv420p(progressive), 1440x3040 [SAR 1:1 DAR 9:19], 30 fps, 30 tbr, 1200k tbn, 60 tbc
Output #0, mp4, to 'full-video.mp4':
  Metadata:
    encoder         : Lavf58.76.100
  Stream #0:0: Video: h264 (High) (avc1 / 0x31637661), yuv420p(progressive), 1440x3040 [SAR 1:1 DAR 9:19], q=2-31, 30 fps, 30 tbr, 90k tbn, 1200k tbc
Stream mapping:
  Stream #0:0 -> #0:0 (copy)
Press [q] to stop, [?] for help
[mp4 @ 0x7f9381864600] Timestamps are unset in a packet for stream 0. This is deprecated and will stop working in the future. Fix your code to set the timestamps properly
[mp4 @ 0x7f9381864600] pts has no value
[mp4 @ 0x7f9381864600] pts has no value0kB time=-00:00:00.03 bitrate=N/A speed=N/A    
    Last message repeated 17981 times
frame=17983 fps=0.0 q=-1.0 Lsize=   52976kB time=00:09:59.36 bitrate= 724.1kbits/s speed=1.33e+03x    
video:52893kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB muxing overhead: 0.156232%


  


  But the most worrisome error for me is from stitching together all interim .mp4 files into one :

  


  [mov,mp4,m4a,3gp,3g2,mj2 @ 0x7f9ff2010e00] Auto-inserting h264_mp4toannexb bitstream filter
Input #0, concat, from 'ffconcat-h264.txt':
  Duration: N/A, bitrate: 1082 kb/s
  Stream #0:0(und): Video: h264 (High) (avc1 / 0x31637661), yuv420p, 1440x3040 [SAR 1:1 DAR 9:19], 1082 kb/s, 30 fps, 30 tbr, 90k tbn, 60 tbc
    Metadata:
      handler_name    : VideoHandler
      vendor_id       : [0][0][0][0]
Output #0, mp4, to 'video-out.mp4':
  Metadata:
    encoder         : Lavf58.76.100
  Stream #0:0(und): Video: h264 (High) (avc1 / 0x31637661), yuv420p, 1440x3040 [SAR 1:1 DAR 9:19], q=2-31, 1082 kb/s, 30 fps, 30 tbr, 90k tbn, 90k tbc
    Metadata:
      handler_name    : VideoHandler
      vendor_id       : [0][0][0][0]
Stream mapping:
  Stream #0:0 -> #0:0 (copy)
Press [q] to stop, [?] for help
[mov,mp4,m4a,3gp,3g2,mj2 @ 0x7f9fe1009c00] Auto-inserting h264_mp4toannexb bitstream filter
[mp4 @ 0x7f9ff2023400] Non-monotonous DTS in output stream 0:0; previous: 53954460, current: 53954460; changing to 53954461. This may result in incorrect timestamps in the output file.
[mov,mp4,m4a,3gp,3g2,mj2 @ 0x7f9fd1008a00] Auto-inserting h264_mp4toannexb bitstream filter
[mp4 @ 0x7f9ff2023400] Non-monotonous DTS in output stream 0:0; previous: 107900521, current: 107874150; changing to 107900522. This may result in incorrect timestamps in the output file.
[mp4 @ 0x7f9ff2023400] Non-monotonous DTS in output stream 0:0; previous: 107900522, current: 107886150; changing to 107900523. This may result in incorrect timestamps in the output file.
frame=36070 fps=0.0 q=-1.0 Lsize=  132464kB time=00:27:54.26 bitrate= 648.1kbits/s speed=6.54e+03x    
video:132296kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB muxing overhead: 0.126409%


  


  I'm not sure how to deal with those non-monotonous DTS errors, and no matter what I try, nothing budges. I analyzed the interim .mp4 files using ffprobe -show_frames and found that the last frame of each interim .mp4 does not have DTS, while previous frames do. E.g. :

  


  ...
[FRAME]
media_type=video
stream_index=0
key_frame=0
pkt_pts=53942461
pkt_pts_time=599.360678
pkt_dts=53942461
pkt_dts_time=599.360678
best_effort_timestamp=53942461
best_effort_timestamp_time=599.360678
pkt_duration=3600
pkt_duration_time=0.040000
pkt_pos=54161377
pkt_size=1034
width=1440
height=3040
pix_fmt=yuv420p
sample_aspect_ratio=1:1
pict_type=B
coded_picture_number=17982
display_picture_number=0
interlaced_frame=0
top_field_first=0
repeat_pict=0
color_range=unknown
color_space=unknown
color_primaries=unknown
color_transfer=unknown
chroma_location=left
[/FRAME]
[FRAME]
media_type=video
stream_index=0
key_frame=0
pkt_pts=53927461
pkt_pts_time=599.194011
pkt_dts=N/A
pkt_dts_time=N/A
best_effort_timestamp=53927461
...


  


  My guess is that as concat demuxer reads in (or elsewhere in ffmpeg's conversion pipeline), for the last frame it sees no DTS set, and produces a virtual value equal to the last seen. Then further in pipeline it consumes this input, sees that DTS value is being repeated, issues a warning and offsets it with increment by one, which might be somewhat nonsensical/unrealistic timing value.

  


  I tried using -fflags +genpts as suggested in this SO answer, but that doesn't change anything.

  


  Per yet other posts suggesting issue being with incompatible tbn and tbc values and possible timebase issues, I tried adding -time_base 1:90000 and -enc_time_base 1:90000 and -copytb 1 and nothing budges. The -video_track_timescale 90000 is there b/c it helped reduce those DTS warnings from 100s down to 3, but doesn't eliminate them all.

  


  Question

  


  What is missing and how can I get ffmpeg to perform conversions without these warnings, to be sure it produces proper, well-formed output ?